---
title: Archive · Local AI Release Board
url: https://recentmodel.com/archive/
updated: 2026-10-24T23:23:29Z
---

# Archive

237 models that left the board (never deleted). Data: [https://recentmodel.com/api/archive.json](https://recentmodel.com/api/archive.json).

- [owensong/Inflect-Mini-v3](https://huggingface.co/owensong/Inflect-Mini-v3) · 2026-10-23
  Mini is the strongest current candidate when audio quality matters most.
- [mingqiancha233/Socrates-Image-1.0](https://huggingface.co/mingqiancha233/Socrates-Image-1.0) · 2026-10-22 · qwen-research
  Socrates-Image-1.0 is a self-contained dense text-to-image model with recurrent visual-semantic feedback. License: qwen-research.
- [FrostWolfAI/Frost-Cyber-Mini-1.0](https://huggingface.co/FrostWolfAI/Frost-Cyber-Mini-1.0) · 2026-10-22 · cc-by-nc-sa-4.0
  Frost Cyber Mini 1.0 is a LoRA fine‑tune of the abliterated (uncensored) Qwen3.5‑4B hybrid model, tuned in two phases on a broad cybersecurity corpus plus offensive / OSINT / threat‑intelligence data. 4.2B params. License: cc-by-nc-sa-4.0.
  Runs on: need Q4 4.2 GB / Q8 6.0 GB / FP16 9.9 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [Remek/basal-omni-pico](https://huggingface.co/Remek/basal-omni-pico) · 2026-10-22 · apache-2.0
  basal-omni-pico is our smallest typed-decision model for Polish and English. Dense, 752M params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.9 GB / Q8 2.2 GB / FP16 2.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [OrisTeam/Aster-0.5B](https://huggingface.co/OrisTeam/Aster-0.5B) · 2026-10-22 · cc-by-sa-4.0
  Aster-0.5B is a 520M-parameter Polish decoder-only causal language model, pretrained from scratch as part of the Aster model family. Dense, 521M params. Context: 4K tokens (config). License: cc-by-sa-4.0.
  Runs on: need Q4 1.4 GB / Q8 1.7 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Phase-Technologies/FORGE-500M-Chat](https://huggingface.co/Phase-Technologies/FORGE-500M-Chat) · 2026-10-22 · apache-2.0
  FORGE-500M-Chat is a hybrid sub-billion parameter conversational language model designed from first principles by Soham Pal. 500M params. License: apache-2.0. Vendor-reported: "FORGE-500M-Chat is a state-of-the-art hybrid sub-billion parameter conversational language model designed from first principles by Soham Pal."
  Runs on: need Q4 1.8 GB / Q8 2.0 GB / FP16 2.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [erayyapagci/balbal-v1](https://huggingface.co/erayyapagci/balbal-v1) · 2026-10-21 · balbal-community-license-1.0
  Negative = balbal-v1 is better; an interval that excludes 0 is a significant difference. License: balbal-community-license-1.0. Vendor-reported: "On read speech (FLEURS) it ties Duyu-small, beats MihuAI and Whisper base, and is behind the much larger Whisper large-v3-turbo and omniASR LLM 7B." Vendor-reported: "Where both models are good (FLEURS, general dev set), steering beats either one alone."
- [ReadyArt/Wicked-Tutu-26B-A4B-Final](https://huggingface.co/ReadyArt/Wicked-Tutu-26B-A4B-Final) · 2026-10-21 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 26B total, 4B active (128 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 21.6 GB / Q8 30.9 GB / FP16 54.6 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Teriminal-RSI/Terminal-RSI-27B](https://huggingface.co/Teriminal-RSI/Terminal-RSI-27B) · 2026-10-21 · apache-2.0
  Terminal-RSI-27B is a Qwen3.8-27B-based terminal-agent model released by the Terminal-RSI project. Dense, 28B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 19.8 GB / Q8 32.6 GB / FP16 58.7 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [ReadyArt/Swish-31B-Final](https://huggingface.co/ReadyArt/Swish-31B-Final) · 2026-10-21 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Dense, 31B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 28.5 GB / Q8 43.1 GB / FP16 71.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [opencerebral/Boris-2-Preview-2](https://huggingface.co/opencerebral/Boris-2-Preview-2) · 2026-10-21 · apache-2.0
  This preview is the main run's 100B-token checkpoint with a short learning-rate decay (2.5B tokens) applied, the same decay length as Boris-2-Preview-1. Dense, 144M params. Context: 2K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.6 GB / Q8 1.7 GB / FP16 1.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ReadyArt/Omega-Arsenal-31B-Final](https://huggingface.co/ReadyArt/Omega-Arsenal-31B-Final) · 2026-10-21 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Dense, 31B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 28.5 GB / Q8 43.1 GB / FP16 71.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [udgamlabs/Laghu](https://huggingface.co/udgamlabs/Laghu) · 2026-10-21 · apache-2.0
  Laghu is a 152M-parameter English chat model trained from scratch by Udgam Labs for $146 of pretraining compute. Dense, 152M params. Context: 16K tokens (config). License: apache-2.0. Vendor-reported: "We fixed these categories, rivals and statistics in advance and publish every result as measured." Vendor-reported: "Compute gap: Falcon-H1-Tiny-90M-Instruct, which beats it everywhere we measured, trained on 800B tokens with a preference-tuning stage, about 12× Laghu's training compute (from TII's published description; our estimate)."
  Runs on: need Q4 1.2 GB / Q8 1.3 GB / FP16 1.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [KerzgrAI/Senryu-2-1B](https://huggingface.co/KerzgrAI/Senryu-2-1B) · 2026-10-21 · apache-2.0
  Senryu-2-1B is the first model on the Senryu-2 architecture and the first release from KerzgrAI. Dense, 1.1B params. Context: 32K tokens (config). License: apache-2.0.
  Runs on: need Q4 2.2 GB / Q8 2.7 GB / FP16 3.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [XtrCom/MINICPM5-2B-SFT-Abiliterated](https://huggingface.co/XtrCom/MINICPM5-2B-SFT-Abiliterated) · 2026-10-21 · apache-2.0
  Text generation model from XtrCom (HF pipeline tag; the card has no one-line description). Dense, 2.5B params. Context: 128K tokens (config). License: apache-2.0.
  Runs on: need Q4 3.0 GB / Q8 4.1 GB / FP16 6.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ATH-MaaS/OvisOCR3](https://huggingface.co/ATH-MaaS/OvisOCR3) · 2026-10-21 · apache-2.0
  OvisOCR3 is a unified 0.8B multimodal model for OCR and document understanding. Dense, 853M params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.9 GB / Q8 2.3 GB / FP16 3.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [caikybaldo999/zyi-256-flash](https://huggingface.co/caikybaldo999/zyi-256-flash) · 2026-10-21 · apache-2.0
  Text to image model from caikybaldo999 (HF pipeline tag; the card has no one-line description). License: apache-2.0.
- [TencentARC/Looking-Ahead](https://huggingface.co/TencentARC/Looking-Ahead) · 2026-10-21 · apache-2.0
  Image to video model from TencentARC (HF pipeline tag; the card has no one-line description). License: apache-2.0.
- [WAS/Minimax-h3_Singularity_v2_pruned](https://huggingface.co/WAS/Minimax-h3_Singularity_v2_pruned) · 2026-10-21 · minimax-h3-community-license-agreement
  Base model: MiniMax H3 by MiniMax. License: minimax-h3-community-license-agreement.
- [dongfangshuo/Dongfangshuo-Qwen3.8-27B-NSFW-v4](https://huggingface.co/dongfangshuo/Dongfangshuo-Qwen3.8-27B-NSFW-v4) · 2026-10-21 · qwen-community-1.0
  Image text to text model from dongfangshuo (HF pipeline tag; the card has no one-line description). Dense, 27B params. Context: 256K tokens (config). License: qwen-community-1.0.
  Runs on: need Q4 19.6 GB / Q8 31.7 GB / FP16 60.4 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Plans11/AGA.Hybrid.Mind.11m](https://huggingface.co/Plans11/AGA.Hybrid.Mind.11m) · 2026-10-21 · other
  Merged from Plans11/AGA-Hybrid-Mind-11m-code on 2026-10-23; that repo has since been removed, so this repo is the one and only home. 11M params. License: other.
  Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [macrocosm-os/iota-decide-4b](https://huggingface.co/macrocosm-os/iota-decide-4b) · 2026-10-21 · apache-2.0
  Image text to text model from macrocosm-os (HF pipeline tag; the card has no one-line description). Dense, 4.5B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 4.8 GB / Q8 6.6 GB / FP16 11.3 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [TheSanchezian/Ternary-Bonsai-2-27B-int2-ov](https://huggingface.co/TheSanchezian/Ternary-Bonsai-2-27B-int2-ov) · 2026-10-21 · apache-2.0
  Bonsai 2 is a ternary version of Qwen3.8 27B. Dense, 27B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 19.3 GB / Q8 31.8 GB / FP16 57.1 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [ReadyArt/Wicked-Tutu-31B-Final](https://huggingface.co/ReadyArt/Wicked-Tutu-31B-Final) · 2026-10-20 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Dense, 31B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 27.8 GB / Q8 42.5 GB / FP16 71.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [realrebelai/MOSS-Imagine_ComfyUI](https://huggingface.co/realrebelai/MOSS-Imagine_ComfyUI) · 2026-10-20 · other
  MOSS-Imagine is an experimental, adaptation-based finetune built upon MOSS-Image SFT, using a curated visual dataset generated with Grok Imagine v0.9 and v1.0. License: other.
- [labqoat/qev-0.8b](https://huggingface.co/labqoat/qev-0.8b) · 2026-10-20 · apache-2.0
  Image text to text model from labqoat (HF pipeline tag; the card has no one-line description). Dense, 853M params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.9 GB / Q8 2.3 GB / FP16 2.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Nianchen/IndustryForge-27B](https://huggingface.co/Nianchen/IndustryForge-27B) · 2026-10-20 · apache-2.0
  IndustryForge-27B is a domain-enhanced multimodal foundation model for industrial CAD. Dense, 27B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 19.6 GB / Q8 32.3 GB / FP16 57.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [ACERobotics/ACE-Brain-0.6-9B](https://huggingface.co/ACERobotics/ACE-Brain-0.6-9B) · 2026-10-20 · mit
  ACE-Brain-0.6-9B is a 9B vision-language model that strengthens spatial grounding for physical foundation models. Dense, 9.4B params. Context: 256K tokens (config). License: mit.
  Runs on: need Q4 7.7 GB / Q8 12.3 GB / FP16 20.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [SubMaroon/Kitchoon-V3-26B-A4B](https://huggingface.co/SubMaroon/Kitchoon-V3-26B-A4B) · 2026-10-20 · gemma
  The Kitchoon series is designed as an RP- and ERP-oriented model, focusing on character development and featuring a neutral writing style. Mixture-of-experts, 26B total, 4B active (128 experts, 8 per token). Context: 256K tokens (config). License: gemma.
  Runs on: need Q4 19.8 GB / Q8 29.9 GB / FP16 54.6 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [dongfangshuo/Dongfangshuo-Qwen3.8-Flash-Next-NSFW](https://huggingface.co/dongfangshuo/Dongfangshuo-Qwen3.8-Flash-Next-NSFW) · 2026-10-20 · qwen-community-1.0
  Image text to text model from dongfangshuo (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 180B total (512 experts, 10 per token). Context: 256K tokens (config). License: qwen-community-1.0.
  Runs on: need Q4 109.8 GB / Q8 193.0 GB / FP16 361.8 GB; smallest fit Q4: 1× DGX Spark (128GB), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB
- [KRAFTON/Raon-1-mini-preview-dflash2](https://huggingface.co/KRAFTON/Raon-1-mini-preview-dflash2) · 2026-10-20 · apache-2.0
  Raon-1 mini is a 30B-A4B Mixture-of-Experts (MoE) language model, designed and trained from scratch by KRAFTON AI. Dense, 678M params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.5 GB / Q8 1.8 GB / FP16 2.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [kasparovabi/drama-3-tr-speaker](https://huggingface.co/kasparovabi/drama-3-tr-speaker) · 2026-10-20 · fish-audio-community-license
  Drama 3 is a 40-layer hybrid model (30 Gated DeltaNet layers, 10 full-attention layers) with 256 routed FP8 experts per layer, 8 active per token, plus a shared expert. License: fish-audio-community-license. Vendor-reported: "In our ablation it beats plain function matching at every scale."
- [ReadyArt/Wicked-Tutu-31B-v2.1](https://huggingface.co/ReadyArt/Wicked-Tutu-31B-v2.1) · 2026-10-19 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Dense, 31B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 27.9 GB / Q8 42.3 GB / FP16 71.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [openbmb/tiny-processor-minicpm-ocr](https://huggingface.co/openbmb/tiny-processor-minicpm-ocr) · 2026-10-19
- [ReadyArt/Wicked-Tutu-31B-v2.0](https://huggingface.co/ReadyArt/Wicked-Tutu-31B-v2.0) · 2026-10-19 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Dense, 31B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 28.5 GB / Q8 43.1 GB / FP16 71.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [baselabs/mammoth-1](https://huggingface.co/baselabs/mammoth-1) · 2026-10-19 · mit
  mammoth-1 is a 27B parameter decision model post-trained from Qwen3.5-27B by Base Labs at Baseten. Dense, 27B params. Context: 256K tokens (config). License: mit.
  Runs on: need Q4 19.6 GB / Q8 32.3 GB / FP16 57.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [robertbschuette/Cobijada_Minimax-H3_Hybrid_Pruned_ComfyUI](https://huggingface.co/robertbschuette/Cobijada_Minimax-H3_Hybrid_Pruned_ComfyUI) · 2026-10-19 · apache-2.0
  Image to video model from robertbschuette (HF pipeline tag; the card has no one-line description). License: apache-2.0.
- [incoai/Qwen3.8-Flash-Next-DFlash2](https://huggingface.co/incoai/Qwen3.8-Flash-Next-DFlash2) · 2026-10-19 · apache-2.0
  DFlash 2 is a block-diffusion drafter for speculative decoding. Dense, 603M params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.6 GB / Q8 1.9 GB / FP16 2.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [SeerRay-Lab/Xiaomi-Lens-0](https://huggingface.co/SeerRay-Lab/Xiaomi-Lens-0) · 2026-10-19 · apache-2.0
  Xiaomi-Lens-0 is a 4B multimodal language model built on Qwen3.5-4B. Dense, 4.6B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 4.9 GB / Q8 7.0 GB / FP16 11.3 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [promptinjections/Qwen3.8-27B-SecOPD-lr6e-5](https://huggingface.co/promptinjections/Qwen3.8-27B-SecOPD-lr6e-5) · 2026-10-19 · apache-2.0
  Image text to text model from promptinjections (HF pipeline tag; the card has no one-line description). Dense, 28B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 19.8 GB / Q8 32.6 GB / FP16 58.7 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [malcolmrey/moss-serenity](https://huggingface.co/malcolmrey/moss-serenity) · 2026-10-19 · apache-2.0
  MOSS Serenity is a full finetune of MOSS-Image Base that teaches the model a set of people by name, while keeping what the base model already knows. License: apache-2.0.
- [xtools-at/Qwen3.5-9B-Qwensploit-CyberSecurity](https://huggingface.co/xtools-at/Qwen3.5-9B-Qwensploit-CyberSecurity) · 2026-10-19 · apache-2.0
  Qwensploit is a cybersecurity-focused fine-tune of a Qwen3.5-9B model merge, designed for security reasoning, vulnerability analysis, secure coding, red-team knowledge, and technical question answering. Dense, 9.4B params. Context: 1M tokens (config). License: apache-2.0.
  Runs on: need Q4 7.7 GB / Q8 11.6 GB / FP16 20.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [TenStrip/MOSSkr2](https://huggingface.co/TenStrip/MOSSkr2) · 2026-10-19 · apache-2.0
  License: apache-2.0.
- [qft-333/V-JEPA-Hand](https://huggingface.co/qft-333/V-JEPA-Hand) · 2026-10-19
  Other model from qft-333 (HF pipeline tag; the card has no one-line description). 607M params.
  Runs on: need Q4 1.9 GB / Q8 2.1 GB / FP16 2.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [gradium/phonon-40m-multilingual](https://huggingface.co/gradium/phonon-40m-multilingual) · 2026-10-19 · cc-by-nc-4.0
  phonon-40m is a text-to-speech model from the Phonon family. 51M params. License: cc-by-nc-4.0.
  Runs on: need Q4 1.5 GB / Q8 1.6 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [XHToken/Spark-X2.6-4B-Base](https://huggingface.co/XHToken/Spark-X2.6-4B-Base) · 2026-10-19 · apache-2.0
  The Spark-X2.6 model series is licensed under the Apache 2.0 License. Dense, 4.1B params. Context: 1M tokens (config). License: apache-2.0. Vendor-reported: "They achieve state-of-the-art performance among similarly sized open-source models, matching or even surpassing models with approximately 9B parameters."
  Runs on: need Q4 4.7 GB / Q8 6.6 GB / FP16 10.4 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [gradium/phonon-voices](https://huggingface.co/gradium/phonon-voices) · 2026-10-19 · apache-2.0
  Each voice is a .safetensors embedding that Phonon uses directly, with no cloning (format ba7a5a45@300: one tensor speakerwavs, 1×512×125, float32). License: apache-2.0.
- [gradium/phonon-40m-fr](https://huggingface.co/gradium/phonon-40m-fr) · 2026-10-19 · cc-by-nc-4.0
  phonon-40m is a text-to-speech model from the Phonon family. 51M params. License: cc-by-nc-4.0.
  Runs on: need Q4 1.5 GB / Q8 1.6 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [gradium/phonon-40m-es](https://huggingface.co/gradium/phonon-40m-es) · 2026-10-19 · cc-by-nc-4.0
  phonon-40m is a text-to-speech model from the Phonon family. 51M params. License: cc-by-nc-4.0.
  Runs on: need Q4 1.5 GB / Q8 1.6 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [gradium/phonon-40m-de](https://huggingface.co/gradium/phonon-40m-de) · 2026-10-19 · cc-by-nc-4.0
  phonon-40m is a text-to-speech model from the Phonon family. 51M params. License: cc-by-nc-4.0.
  Runs on: need Q4 1.5 GB / Q8 1.6 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [gradium/phonon-40m-pt](https://huggingface.co/gradium/phonon-40m-pt) · 2026-10-19 · cc-by-nc-4.0
  phonon-40m is a text-to-speech model from the Phonon family. 51M params. License: cc-by-nc-4.0.
  Runs on: need Q4 1.5 GB / Q8 1.6 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [incoai/Ornith-1.5-122B-DFlash2](https://huggingface.co/incoai/Ornith-1.5-122B-DFlash2) · 2026-10-19 · apache-2.0
  DFlash 2 is a block-diffusion drafter for speculative decoding. Dense, 930M params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.8 GB / Q8 2.2 GB / FP16 3.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [baselabs/elephant-1](https://huggingface.co/baselabs/elephant-1) · 2026-10-19 · mit
  elephant-1 is a 9B parameter decision model post-trained from Qwen3.5-9B by Base Labs at Baseten. Dense, 9.4B params. Context: 256K tokens (config). License: mit. Vendor-reported: "On the public split, elephant-1 narrowly trails mammoth-1 in every area and beats Jev 1.14 in all five (knowledge & reasoning, language, tools & automation, retrieval & classification, arts & human taste)."
  Runs on: need Q4 7.7 GB / Q8 12.3 GB / FP16 20.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [agbalu/Jugurtha-4B](https://huggingface.co/agbalu/Jugurtha-4B) · 2026-10-18 · apache-2.0
  Text generation model from agbalu (HF pipeline tag; the card has no one-line description). Dense, 4.7B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 4.9 GB / Q8 6.7 GB / FP16 11.4 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [AlgoDriveAI/Inferno-Exp-1.7](https://huggingface.co/AlgoDriveAI/Inferno-Exp-1.7) · 2026-10-18 · apache-2.0
  Inferno-Exp-1.7 is AlgoDriveAI's 393.6M-parameter, math-first decoder-only base language model. Dense, 394M params. Context: 4K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.5 GB / Q8 1.7 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ReadyArt/Wicked-Tutu-26B-A4B](https://huggingface.co/ReadyArt/Wicked-Tutu-26B-A4B) · 2026-10-18 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 26B total, 4B active (128 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 21.6 GB / Q8 30.9 GB / FP16 54.6 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [ReadyArt/Wicked-Tutu-26B-A4B-v0.26](https://huggingface.co/ReadyArt/Wicked-Tutu-26B-A4B-v0.26) · 2026-10-18 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 26B total, 4B active. License: apache-2.0.
  Runs on: need Q4 20.0 GB / Q8 30.1 GB / FP16 54.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [lodestones/MOSS-Image-Base-8-steps-dmd](https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd) · 2026-10-18 · apache-2.0
  5.1B params. License: apache-2.0.
  Runs on: need Q4 4.5 GB / Q8 6.9 GB / FP16 11.7 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [congruency/MOSS-Image-Turbo-MFD](https://huggingface.co/congruency/MOSS-Image-Turbo-MFD) · 2026-10-18 · apache-2.0
  Text to image model from congruency (HF pipeline tag; the card has no one-line description). License: apache-2.0.
- [berryber09/MiniMax-H3-FastH3-VSA-4step-pruned-w4a8](https://huggingface.co/berryber09/MiniMax-H3-FastH3-VSA-4step-pruned-w4a8) · 2026-10-18 · minimax-h3-community-license
  Image to video model from berryber09 (HF pipeline tag; the card has no one-line description). License: minimax-h3-community-license.
- [RuVI-AI/Pragya-EXP2](https://huggingface.co/RuVI-AI/Pragya-EXP2) · 2026-10-18 · apache-2.0
  Text generation model from RuVI-AI (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 29M total (8 experts, 2 per token). License: apache-2.0.
  Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ropedia-ai/Ropedia-W0](https://huggingface.co/ropedia-ai/Ropedia-W0) · 2026-10-18 · cc-by-nc-4.0
  Ropedia W0 is an egocentric world foundation model that learns loco-manipulation from 10k hours of human experience and transfers it across embodiments, from robot arms to humanoids and from grippers to dexterous hands, through one shared kinematic interface of camera poses and projected skeletons. License: cc-by-nc-4.0.
- [migtissera/Tess-4.5-27B](https://huggingface.co/migtissera/Tess-4.5-27B) · 2026-10-18 · apache-2.0
  Tess-4.5 is built to ask before destructive actions and to report failures plainly, but no model is a guarantee. Dense, 28B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 21.7 GB / Q8 34.9 GB / FP16 57.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [fyxfelixfu/consistentDMD](https://huggingface.co/fyxfelixfu/consistentDMD) · 2026-10-18 · cc-by-nc-sa-4.0
  See consistentdmd-ldit-best.pt.json for model and EMA metadata, consistentdmd-ldit-best.pt.sha256 for its checksum, and SHA256SUMS for all release files. License: cc-by-nc-sa-4.0.
- [SalahALHaismawi/Nuqta-OCR-0.8B](https://huggingface.co/SalahALHaismawi/Nuqta-OCR-0.8B) · 2026-10-18 · apache-2.0
  Nuqta-OCR 0.8B is an open-weights Arabic OCR model. Dense, 890M params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 2.0 GB / Q8 2.3 GB / FP16 3.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [SalahALHaismawi/Nuqta-OCR-4B](https://huggingface.co/SalahALHaismawi/Nuqta-OCR-4B) · 2026-10-18 · apache-2.0
  Nuqta-OCR 4B is an open-weights Arabic OCR model. Dense, 4.8B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 5.0 GB / Q8 6.9 GB / FP16 11.8 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [Blackfrost-AI/NVIDIA-Nemotron-3.5-Bolt-74B-DERISKED-BF16](https://huggingface.co/Blackfrost-AI/NVIDIA-Nemotron-3.5-Bolt-74B-DERISKED-BF16) · 2026-10-17 · nvidia-nemotron-open-model-license-inherited
  Text generation model from Blackfrost-AI (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 75B total, 9B active (512 experts). Context: 256K tokens (config). License: nvidia-nemotron-open-model-license-inherited.
  Runs on: need Q4 53.8 GB / Q8 86.3 GB / FP16 157.2 GB; smallest fit Q4: Mac 96GB, Q8: 96GB GPU (RTX Pro 6000 Blackwell), FP16: 2× DGX Spark (256GB)
- [Team-CorerepubliX/Gureum_1.0_neo](https://huggingface.co/Team-CorerepubliX/Gureum_1.0_neo) · 2026-10-17 · cc-by-nc-nd-4.0
  Text generation model from Team-CorerepubliX (HF pipeline tag; the card has no one-line description). Dense, 2.3B params. Context: 32K tokens (config). License: cc-by-nc-nd-4.0.
  Runs on: need Q4 4.0 GB / Q8 5.1 GB / FP16 7.2 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [ReadyArt/Wicked-Tutu-31B-v0.25](https://huggingface.co/ReadyArt/Wicked-Tutu-31B-v0.25) · 2026-10-17 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). 31B params. License: apache-2.0.
  Runs on: need Q4 23.1 GB / Q8 37.7 GB / FP16 66.2 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [ReadyArt/Wicked-Tutu-31B-v0.2](https://huggingface.co/ReadyArt/Wicked-Tutu-31B-v0.2) · 2026-10-17 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). 31B params. License: apache-2.0.
  Runs on: need Q4 23.1 GB / Q8 36.9 GB / FP16 66.2 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Gen-Verse/DiscoveryRSI-27B](https://huggingface.co/Gen-Verse/DiscoveryRSI-27B) · 2026-10-17 · apache-2.0
  DiscoveryRSI-27B is an instruction-finetuned 27B language model intended for research ideation and evidence-grounded planning. Dense, 28B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 19.8 GB / Q8 32.6 GB / FP16 58.7 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Blockway/Agens-Volundr-32B-v1](https://huggingface.co/Blockway/Agens-Volundr-32B-v1) · 2026-10-17 · apache-2.0
  Agens Volundr 32B is Blockway's open-weights model for coding, reasoning and agent work — the model behind Agens. v1 is the full release the Preview (2026-10-05) announced: the same architecture, continued pre-training run to its end, and post-training built for long agent sessions. Dense, 33B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 23.0 GB / Q8 38.1 GB / FP16 68.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [dx8152/Qwen-Image-2.1-Font](https://huggingface.co/dx8152/Qwen-Image-2.1-Font) · 2026-10-17 · apache-2.0
  License: apache-2.0.
- [Arko007/Zenyx-100M-Base](https://huggingface.co/Arko007/Zenyx-100M-Base) · 2026-10-17 · apache-2.0
  Base model, no fine-tuning, 0-shot, all test items. 100M params. License: apache-2.0. Vendor-reported: "Zenyx-100M beats BetterGPT in 8 of 22 domains (GitHub, YouTube Subtitles, EuroParl, Wikipedia (en), StackExchange, DM Mathematics, Ubuntu IRC, PhilPapers) and is 1 to 6% behind in the other 14 (overall 0.9% behind)." Vendor-reported: "Reading the memory in the cached decoder is exact (logits agree with the full forward, top-1 agreement 100%, greedy decode identical on 16 of 16 tokens)."
  Runs on: need Q4 1.6 GB / Q8 1.6 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [MemTensor/MemT-1.0-Flash](https://huggingface.co/MemTensor/MemT-1.0-Flash) · 2026-10-17 · mit
  MemT-1.0 is a new generation of memory foundation models that makes persistent, dynamically evolving memory an intrinsic part of model computation. Mixture-of-experts, 288B total (256 experts, 6 per token). Context: 1M tokens (config). License: mit. Vendor-reported: "Across 11 memory benchmarks, MemT-1.0-Flash and MemT-1.0-Nano rank first and second among the evaluated parametric-memory methods, respectively."
  Runs on: need Q4 174.5 GB / Q8 307.6 GB / FP16 577.6 GB; smallest fit Q4: 2× DGX Spark (256GB), Q8: Mac 512GB
- [FastVideo/FastH3-QAD-v2](https://huggingface.co/FastVideo/FastH3-QAD-v2) · 2026-10-17 · other
  FastH3-QAD-v2 is an 8-step video-and-audio model with NVFP4 weights, trained with quantization-aware distillation (QAD). 14B params. License: other.
  Runs on: need Q4 10.5 GB / Q8 16.9 GB / FP16 29.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 24GB GPU (3090 / 4090), FP16: 32GB GPU (5090)
- [ReadyArt/Broken-Tutu-Unleashed-31B-v0.1](https://huggingface.co/ReadyArt/Broken-Tutu-Unleashed-31B-v0.1) · 2026-10-16 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). 31B params. License: apache-2.0.
  Runs on: need Q4 23.1 GB / Q8 37.7 GB / FP16 66.2 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [XingChen-AGI/Xing4.0-29B-A4B-w8a8](https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B-w8a8) · 2026-10-16 · apache-2.0
  Xing4.0-29B-A4B is a large language model in the Xing series (formerly TeleChat), developed by China Telecom Artificial Intelligence Technology Co., Ltd. Mixture-of-experts, 29B total, 4B active (64 experts, 4 per token). Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 18.8 GB / Q8 32.2 GB / FP16 59.4 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [akhaliq/MOSS-Image-SFT-4-steps](https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps) · 2026-10-16 · apache-2.0
  At 4 steps this model matches the 8-step checkpoint on text accuracy and photographic detail, and clearly beats running the 8-step checkpoint truncated to 4 steps (sharper, better contrast, more accurate Chinese text). 5.1B params. License: apache-2.0. Vendor-reported: "At 4 steps this model matches the 8-step checkpoint on text accuracy and photographic detail, and clearly beats running the 8-step checkpoint truncated to 4 steps (sharper, better contrast, more accurate Chinese text)."
  Runs on: need Q4 4.5 GB / Q8 6.9 GB / FP16 11.7 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [DedeProGames/GPT-U-20M-Instruct](https://huggingface.co/DedeProGames/GPT-U-20M-Instruct) · 2026-10-16 · apache-2.0
  Text generation model from DedeProGames (HF pipeline tag; the card has no one-line description). Dense, 20M params. Context: 1K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.1 GB / Q8 1.1 GB / FP16 1.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [sKT-Ai-Labs/SPEED-1-BASE](https://huggingface.co/sKT-Ai-Labs/SPEED-1-BASE) · 2026-10-16 · mit
  SPEED-1 is a Mixture-of-Experts (MoE) language model with ~19B parameters, built by SKT AI LABS Private Limited, an Indian AI company based in Sidhi, Madhya Pradesh. Mixture-of-experts, 19B total (2 experts, 2 per token). Context: 4K tokens (config). License: mit.
  Runs on: need Q4 14.2 GB / Q8 23.0 GB / FP16 41.0 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [kaupane/Inko](https://huggingface.co/kaupane/Inko) · 2026-10-16
  Inko is a bilingual text-to-image model for paintings and illustrations. 537M params.
  Runs on: need Q4 1.8 GB / Q8 2.1 GB / FP16 2.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [knowledgator/gliclass-v4-edge](https://huggingface.co/knowledgator/gliclass-v4-edge) · 2026-10-16 · apache-2.0
  GLiClass answers the same requests directly: each question is one GLiClass pass in which the state is the text, the question's instructions are the prompt, and every option is a label. 440M params. License: apache-2.0.
  Runs on: need Q4 1.8 GB / Q8 2.0 GB / FP16 2.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ReadyArt/Wicked-Tutu-31B-v0.1](https://huggingface.co/ReadyArt/Wicked-Tutu-31B-v0.1) · 2026-10-16 · apache-2.0
  Text generation model from ReadyArt (HF pipeline tag; the card has no one-line description). 31B params. License: apache-2.0.
  Runs on: need Q4 23.1 GB / Q8 37.7 GB / FP16 66.2 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [brandonmusic/Muse-Spark-Open-816B-A42B-TrellisMX-K3](https://huggingface.co/brandonmusic/Muse-Spark-Open-816B-A42B-TrellisMX-K3) · 2026-10-16 · muse-model-license
  A 3-bit TrellisMX build of meta-models/Muse-Spark-Open-816B-A42B (BF16 base, revision 0fb02478b02565b2ce7fc57e6051c91172774ac9) sized for four 96 GB RTX PRO 6000 Blackwell cards at tensor parallel 4, with the image, video and audio towers kept. 24B params. License: muse-model-license. Vendor-reported: "Intervals are window-block bootstrap (5,000 resamples, BCa) and Wilson for top-1." Vendor-reported: "All 128 windows, without the overlap screen: 0.0333 [0.0283, 0.0399], top-1 94.52% (227,038 text rows)."
  Runs on: need Q4 17.2 GB / Q8 28.1 GB / FP16 50.2 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Thanushttk/pocket-tts-multilingual](https://huggingface.co/Thanushttk/pocket-tts-multilingual) · 2026-10-16 · cc-by-4.0
  Text to speech model from Thanushttk (HF pipeline tag; the card has no one-line description). 110M params. License: cc-by-4.0.
  Runs on: need Q4 1.6 GB / Q8 1.6 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ozgurguler/ahenk-cep-tr](https://huggingface.co/ozgurguler/ahenk-cep-tr) · 2026-10-16 · cc-by-4.0
  Ahenk Cep is a small student of the Ahenk model: 8.6M parameters, 34 MB (fp32), EMA Lightning architecture, one designed female voice (f022, no real person), 48 kHz. License: cc-by-4.0.
- [radiant-lyrics/sylla-600m-v1](https://huggingface.co/radiant-lyrics/sylla-600m-v1) · 2026-10-16 · gpl-3.0
  Sylla and the base model were measured by us on the current Hugging Face release of the dataset, so the comparison is indicative rather than a controlled reproduction. 92.9% of Sylla's word onsets land within 200 ms. Dense, 581M params. License: gpl-3.0.
  Runs on: need Q4 2.1 GB / Q8 2.4 GB / FP16 3.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [MaorB/embeddinggemma-2-asr-hebrew-english-streaming](https://huggingface.co/MaorB/embeddinggemma-2-asr-hebrew-english-streaming) · 2026-10-16 · apache-2.0
  Automatic speech recognition model from MaorB (HF pipeline tag; the card has no one-line description). 99M params. License: apache-2.0.
  Runs on: need Q4 1.6 GB / Q8 1.6 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [naver-ai/DroneCATS-Cosmos3-Edge-2B](https://huggingface.co/naver-ai/DroneCATS-Cosmos3-Edge-2B) · 2026-10-16 · openmdw1.1-license
  Image text to text model from naver-ai (HF pipeline tag; the card has no one-line description). Dense, 2.4B params. Context: 128K tokens (config). License: openmdw1.1-license.
  Runs on: need Q4 3.4 GB / Q8 4.5 GB / FP16 6.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Kijai/MOSS-image_experimental](https://huggingface.co/Kijai/MOSS-image_experimental) · 2026-10-16
- [mindlab-research/MJev-Spatial-4B](https://huggingface.co/mindlab-research/MJev-Spatial-4B) · 2026-10-16
  MJev-Spatial is a multimodal decision model for spatial understanding. 4B params. Context: 4K tokens (config).
  Runs on: need Q4 3.9 GB / Q8 5.7 GB / FP16 9.5 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [Tongyi-MAI/MAI-Gym-9B](https://huggingface.co/Tongyi-MAI/MAI-Gym-9B) · 2026-10-16 · apache-2.0
  Dense, 9.4B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 7.7 GB / Q8 12.1 GB / FP16 20.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [nvidia/Muse-Spark-Open-816B-A42B-DFlash2](https://huggingface.co/nvidia/Muse-Spark-Open-816B-A42B-DFlash2) · 2026-10-16 · nvidia-open-model-agreement-license
  Image text to text model from nvidia (HF pipeline tag; the card has no one-line description). Dense, 4.4B params. Context: 1M tokens (config). License: nvidia-open-model-agreement-license.
  Runs on: need Q4 3.8 GB / Q8 5.9 GB / FP16 10.0 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [AlgoDriveAI/Inferno-Exp-1.6](https://huggingface.co/AlgoDriveAI/Inferno-Exp-1.6) · 2026-10-15 · apache-2.0
  Inferno-Exp-1.6 is a 393.6M-parameter decoder-only base language model trained from scratch by AlgoDriveAI on about 4B tokens. Dense, 394M params. Context: 2K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.5 GB / Q8 1.7 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [huggingwadhah/lowendgirlfriend](https://huggingface.co/huggingwadhah/lowendgirlfriend) · 2026-10-15 · mixed-see-license-section
  Text generation model from huggingwadhah (HF pipeline tag; the card has no one-line description). Dense, 52M params. Context: 16K tokens (config). License: mixed-see-license-section.
  Runs on: need Q4 1.1 GB / Q8 1.2 GB / FP16 1.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [public-release-bp/BudgetPix](https://huggingface.co/public-release-bp/BudgetPix) · 2026-10-15 · mit
  Text to image model from public-release-bp (HF pipeline tag; the card has no one-line description). License: mit.
- [RudySen/Krea2-Realism-Checkpoint](https://huggingface.co/RudySen/Krea2-Realism-Checkpoint) · 2026-10-15 · apache-2.0
  License: apache-2.0.
- [MithrilMan/teacher-in-the-loop](https://huggingface.co/MithrilMan/teacher-in-the-loop) · 2026-10-15 · apache-2.0
  Each model speaks one of Kokoro's voices: afheart (English), ifsara and imnicola (Italian). License: apache-2.0.
- [XuanyuLiu/Scene-Builder-Qwen3-VL-8B-Instruct](https://huggingface.co/XuanyuLiu/Scene-Builder-Qwen3-VL-8B-Instruct) · 2026-10-15
  Scene-Builder is the supervised fine-tuning (SFT) model from Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation, built on Qwen3-VL-8B-Instruct. Dense, 8.8B params. Context: 256K tokens (config).
  Runs on: need Q4 7.5 GB / Q8 11.5 GB / FP16 19.7 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [wesjos/mario-html-v2-9B](https://huggingface.co/wesjos/mario-html-v2-9B) · 2026-10-15 · apache-2.0
  Image text to text model from wesjos (HF pipeline tag; the card has no one-line description). Dense, 9.7B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 7.7 GB / Q8 11.6 GB / FP16 21.4 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [j-shen/FlashMap-Qwen3-8B](https://huggingface.co/j-shen/FlashMap-Qwen3-8B) · 2026-10-15 · apache-2.0
  FlashMap is a drafter for lossless speculative decoding. Dense, 500M params. Context: 128K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.7 GB / Q8 1.9 GB / FP16 2.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Remek/basal-omni-1.0-ultra](https://huggingface.co/Remek/basal-omni-1.0-ultra) · 2026-10-15 · apache-2.0
  basal-omni-1.0-ultra (31B) is the largest model of basal-omni, basal's line on Google Gemma 4. Dense, 31B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 28.6 GB / Q8 41.7 GB / FP16 70.5 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [pipecat-ai/phonellm-mini-alpha-2](https://huggingface.co/pipecat-ai/phonellm-mini-alpha-2) · 2026-10-15 · bsd-2-clause
  PhoneLLM Mini Alpha 2 is a full-parameter supervised fine-tune of NVIDIA Nemotron 3 Nano 30B-A3B for phone voice agents with tool calling. Mixture-of-experts, 32B total (128 experts, 6 per token). Context: 256K tokens (config). License: bsd-2-clause.
  Runs on: need Q4 25.9 GB / Q8 35.0 GB / FP16 64.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [XingChen-AGI/Xing4.0-29B-Intent](https://huggingface.co/XingChen-AGI/Xing4.0-29B-Intent) · 2026-10-15 · apache-2.0
  Xing4.0-29B-Intent is a schema-guided model for intent recognition and structured information extraction, developed from Xing4.0-29B-A4B, the Xing semantic large language model from China Telecom Artificial Intelligence Technology Co., Ltd. Mixture-of-experts, 31B total, 4B active (64 experts, 4 per token). Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 20.1 GB / Q8 34.6 GB / FP16 63.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [MemTensor/MemT-1.0-Nano](https://huggingface.co/MemTensor/MemT-1.0-Nano) · 2026-10-15 · apache-2.0
  MemT-1.0 is a new generation of memory foundation models that makes persistent, dynamically evolving memory an intrinsic part of model computation. 35B params. License: apache-2.0. Vendor-reported: "Across 11 memory benchmarks, MemT-1.0-Flash and MemT-1.0-Nano rank first and second among the evaluated parametric-memory methods, respectively."
  Runs on: need Q4 25.1 GB / Q8 41.3 GB / FP16 74.2 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [calon-ai/calon-0.6b](https://huggingface.co/calon-ai/calon-0.6b) · 2026-10-15 · apache-2.0
  CalOn is a 0.6B-parameter encoder for zero-shot entity, relation, and event extraction. 601M params. License: apache-2.0.
  Runs on: need Q4 1.9 GB / Q8 2.1 GB / FP16 2.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [nlpai-lab/KURE-v3](https://huggingface.co/nlpai-lab/KURE-v3) · 2026-10-15 · apache-2.0
  KURE-v3 is a Korean-English bilingual dense retrieval model. Dense, 1.1B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "It outperforms 8B dense models such as comsat-embed-ko-8b-preview and Qwen3-Embedding-8B." Vendor-reported: "Truncated to 1,024 dimensions, KURE-v3 keeps 99% of its full-size score and still outperforms comsat-embed-ko-8b-preview and Qwen3-Embedding-8B (4,096 dimensions)."
  Runs on: need Q4 2.2 GB / Q8 2.7 GB / FP16 3.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [athrael-soju/UltraVBERT](https://huggingface.co/athrael-soju/UltraVBERT) · 2026-10-15 · mit
  UltraVBERT is a 247M-parameter vision-language encoder built as a base for visual document retrievers. Dense, 293M params. Context: 32K tokens (config). License: mit.
  Runs on: need Q4 1.8 GB / Q8 1.9 GB / FP16 2.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [fastino/Fastino-Nemotron-3.5-Super-VL-Healthcare](https://huggingface.co/fastino/Fastino-Nemotron-3.5-Super-VL-Healthcare) · 2026-10-14 · apache-2.0
  Text generation model from fastino (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 124B total (512 experts, 22 per token). Context: 1M tokens (config). License: apache-2.0.
  Runs on: need Q4 85.7 GB / Q8 142.8 GB / FP16 258.7 GB; smallest fit Q4: 96GB GPU (RTX Pro 6000 Blackwell), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB
- [fastino/Fastino-Nemotron-3.5-Super-VL-Finance](https://huggingface.co/fastino/Fastino-Nemotron-3.5-Super-VL-Finance) · 2026-10-14 · apache-2.0
  Text generation model from fastino (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 124B total (512 experts, 22 per token). Context: 1M tokens (config). License: apache-2.0.
  Runs on: need Q4 85.7 GB / Q8 142.8 GB / FP16 258.7 GB; smallest fit Q4: 96GB GPU (RTX Pro 6000 Blackwell), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB
- [startlux-models/ReaDy-Qwen3.8-27B](https://huggingface.co/startlux-models/ReaDy-Qwen3.8-27B) · 2026-10-14
  ReaDy augments a DSpark drafter with position-specific reads from the target model's committed recurrent states, and injects the retrieved context as a residual update to the draft representation. Dense, 1.9B params. Context: 256K tokens (config).
  Runs on: need Q4 2.3 GB / Q8 3.2 GB / FP16 4.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [DedeProGames/Kiyo-230M-Instruct](https://huggingface.co/DedeProGames/Kiyo-230M-Instruct) · 2026-10-14 · apache-2.0
  Kiyo-230M-Instruct is the largest model in this comparison, so it is also the slowest per token on CPU: about 13 tokens/s for a single stream here, or roughly 20 seconds for a 256-token answer. Dense, 230M params. Context: 2K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [zime-ai/dwight-4b](https://huggingface.co/zime-ai/dwight-4b) · 2026-10-13 · apache-2.0
  Dwight is a 4-billion-parameter model from Zime AI, the AI infrastructure platform for sales teams. Dense, 4B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 4.8 GB / Q8 6.5 GB / FP16 10.3 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [domyn/Domyn-Edge-SFT](https://huggingface.co/domyn/Domyn-Edge-SFT) · 2026-10-13 · apache-2.0
  Domyn Edge SFT is the supervised fine-tuned version of Domyn Edge Base, a 4B-parameter dense transformer pre-trained from scratch on 13 trillion tokens, end to end on European public infrastructure. Dense, 3.9B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 4.2 GB / Q8 6.0 GB / FP16 9.7 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [hazyresearch/ncm-checkpoints](https://huggingface.co/hazyresearch/ncm-checkpoints) · 2026-10-13 · mit
  Text generation model from hazyresearch (HF pipeline tag; the card has no one-line description). License: mit.
- [summerMC/Diffusion-Qwen3.8-27B](https://huggingface.co/summerMC/Diffusion-Qwen3.8-27B) · 2026-10-13 · apache-2.0
  All three DiffusionQwen models support opt-in block KV and bidirectional DeltaNet state caching. Dense, 27B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 19.2 GB / Q8 31.7 GB / FP16 56.9 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [knowledgator/gliclass-v4-27b](https://huggingface.co/knowledgator/gliclass-v4-27b) · 2026-10-13 · apache-2.0
  GLiClass answers the same requests directly: each question is one GLiClass pass in which the state is the text, the question's instructions are the prompt, and every option is a label. 27B params. License: apache-2.0.
  Runs on: need Q4 19.5 GB / Q8 32.0 GB / FP16 57.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [lightonai/mSlideOn](https://huggingface.co/lightonai/mSlideOn) · 2026-10-13 · apache-2.0
  SlideOn and mSlideOn reuse these models' retrieval knowledge to learn sparse representations through SAE pretraining followed by supervised fine-tuning. 339M params. License: apache-2.0. Vendor-reported: "Its sparse encoder lifts mLateOn's normalized 128-dimensional token representations to 768 dimensions and applies a residual MLP before selecting the top-32 latent activations and applying ReLU." Vendor-reported: "Reaches 75.79 on MLDR target languages and 61.68 across all 13 languages, outperforming every dense baseline in our comparison."
  Runs on: need Q4 1.7 GB / Q8 1.9 GB / FP16 2.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [AlgoDriveAI/Inferno-Exp-1.5](https://huggingface.co/AlgoDriveAI/Inferno-Exp-1.5) · 2026-10-12 · apache-2.0
  Inferno-Exp-1.5 is a 76.2M-parameter decoder-only base language model trained from scratch by AlgoDriveAI. Dense, 76M params. Context: 2K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.1 GB / Q8 1.2 GB / FP16 1.3 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [mukaddamzaid/amos-1-small](https://huggingface.co/mukaddamzaid/amos-1-small) · 2026-10-12 · apache-2.0
  License: apache-2.0. Vendor-reported: "Rivals run by us on the same 2,000 decisions, except Decider 1, whose number is its makers' own."
- [VAST-AI/EdenLive](https://huggingface.co/VAST-AI/EdenLive) · 2026-10-12 · edenlive-mixed
  Image to video model from VAST-AI (HF pipeline tag; the card has no one-line description). License: edenlive-mixed.
- [switchboardllm/switchboard-router-0.5b](https://huggingface.co/switchboardllm/switchboard-router-0.5b) · 2026-10-12 · apache-2.0
  Switchboard v0.1 is a 0.5B router: it reads a prompt and answers with one line of JSON, the cheapest model tier that can handle it. Dense, 494M params. Context: 32K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.4 GB / Q8 1.6 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Remek/basal-2.0-max](https://huggingface.co/Remek/basal-2.0-max) · 2026-10-12 · apache-2.0
  basal-2.0-max is part of basal-2.0, a family of typed-decision models for Polish and English; it is the most accurate model of the family. Dense, 11B params. Context: 32K tokens (config). License: apache-2.0.
  Runs on: need Q4 9.4 GB / Q8 14.6 GB / FP16 25.0 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 32GB GPU (5090)
- [Prahaladha/hima-asr](https://huggingface.co/Prahaladha/hima-asr) · 2026-10-12 · apache-2.0-and-gemma
  Hima-ASR is a compact open speech recognition model for 11 Indian languages and English: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and English. 587M params. License: apache-2.0-and-gemma. Vendor-reported: "Hima-ASR has the lowest error of the four in Tamil, Telugu, Malayalam and Kannada, and beats Sarvam in Hindi (14.2 vs 16.6)." Vendor-reported: "On the 12-language average it beats Flex and is within 0.2 points of Core (12.84) and Sarvam (12.79); Sarvam is best on English."
  Runs on: need Q4 1.9 GB / Q8 2.1 GB / FP16 2.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [lightonai/SlideOn](https://huggingface.co/lightonai/SlideOn) · 2026-10-12 · apache-2.0
  SlideOn and mSlideOn reuse these models' retrieval knowledge to learn sparse representations through SAE pretraining followed by supervised fine-tuning. 174M params. License: apache-2.0. Vendor-reported: "A learned linear encoder maps each normalized DenseOn token representation to its top-32 latent activations, followed by ReLU."
  Runs on: need Q4 1.6 GB / Q8 1.7 GB / FP16 1.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [thongbuind/SAI_375M](https://huggingface.co/thongbuind/SAI_375M) · 2026-10-12 · apache-2.0
  Text generation model from thongbuind (HF pipeline tag; the card has no one-line description). Dense, 375M params. Context: 10K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.4 GB / Q8 1.6 GB / FP16 2.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [AuroraAI-Research/AuroraIMG-6M](https://huggingface.co/AuroraAI-Research/AuroraIMG-6M) · 2026-10-11 · apache-2.0
  AuroraIMG-6M is a text-to-image model that's just 6.82M parameters, and that number includes everything: the text encoder, the denoiser and the VAE.
  Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [autotrust/JEV-35B](https://huggingface.co/autotrust/JEV-35B) · 2026-10-11 · apache-2.0
  JEV-35B is a decision model only: it does not chat or generate text. Mixture-of-experts, 36B total, 3B active (256 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 23.4 GB / Q8 40.1 GB / FP16 73.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [DICPlay/RusMind-Momentum](https://huggingface.co/DICPlay/RusMind-Momentum) · 2026-10-11 · rusmind-license
  Text generation model from DICPlay (HF pipeline tag; the card has no one-line description). License: rusmind-license.
- [startlux-models/StartLux-Decision-v1.1-27B-Instant](https://huggingface.co/startlux-models/StartLux-Decision-v1.1-27B-Instant) · 2026-10-11 · cc-by-nc-4.0
  StartLux-Decision-v1.1-27B-Instant is a 27B decision model. Dense, 28B params. Context: 256K tokens (config). License: cc-by-nc-4.0.
  Runs on: need Q4 19.8 GB / Q8 32.6 GB / FP16 58.7 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Edisoneh/InfiniSplat-v2](https://huggingface.co/Edisoneh/InfiniSplat-v2) · 2026-10-11 · other
  Image to 3d model from Edisoneh (HF pipeline tag; the card has no one-line description). 1.6B params. License: other.
  Runs on: need Q4 2.5 GB / Q8 3.3 GB / FP16 4.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [DICPlay/RusMind-Vspyshka](https://huggingface.co/DICPlay/RusMind-Vspyshka) · 2026-10-11 · rusmind-license
  Text generation model from DICPlay (HF pipeline tag; the card has no one-line description). License: rusmind-license.
- [dmis-lab/UltraBERT-150M](https://huggingface.co/dmis-lab/UltraBERT-150M) · 2026-10-10 · apache-2.0
  UltraBERT is a family of English bidirectional encoders with a native 32K context that reach on both classification and retrieval, two task families on which previous encoders had to choose a side. Dense, 190M params. License: apache-2.0. Vendor-reported: "UltraBERT is a family of English bidirectional encoders with a native 32K context that reach state of the art on both classification and retrieval, two task families on which previous encoders had to choose a side."
  Runs on: need Q4 1.7 GB / Q8 1.8 GB / FP16 2.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [zoebchhatriwala/Zesyde-150M-Multilingual-Router](https://huggingface.co/zoebchhatriwala/Zesyde-150M-Multilingual-Router) · 2026-10-10 · mit
  Zesyde is a production-ready, non-autoregressive foundation model engineered for ultra-fast, mathematically calibrated zero-shot decisions and agent routing across 1,800+ languages. 150M params. License: mit.
  Runs on: need Q4 1.6 GB / Q8 1.7 GB / FP16 1.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Lalakai/Hunter-1-1.7B](https://huggingface.co/Lalakai/Hunter-1-1.7B) · 2026-10-10 · apache-2.0
  Hunter 1.7B is a compact 1.7B-parameter language model from Lalakai Labs, tuned for defensive security work: incident triage, incident-response question answering, and causal reasoning about attack paths. Dense, 1.7B params. Context: 1M tokens (config). License: apache-2.0.
  Runs on: need Q4 2.6 GB / Q8 3.3 GB / FP16 4.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [294-288-pretrain-interaction/160M-exp](https://huggingface.co/294-288-pretrain-interaction/160M-exp) · 2026-10-10
  Text generation model from 294-288-pretrain-interaction (HF pipeline tag; the card has no one-line description). Dense, 160M params. Context: 2K tokens (config).
  Runs on: need Q4 1.4 GB / Q8 1.5 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [startlux-models/StartLux-Decision-v1.1-27B-AutoThink](https://huggingface.co/startlux-models/StartLux-Decision-v1.1-27B-AutoThink) · 2026-10-10 · cc-by-nc-4.0
  StartLux-Decision-v1.1-27B-AutoThink is a 27B decision model that reasons only where it is unsure. Dense, 27B params. Context: 256K tokens (config). License: cc-by-nc-4.0.
  Runs on: need Q4 19.5 GB / Q8 32.2 GB / FP16 57.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [masterset-ai/PhysicalEye-Decide-35B](https://huggingface.co/masterset-ai/PhysicalEye-Decide-35B) · 2026-10-10 · apache-2.0
  PhysicalEye-Decide-35B is a decision engine for images. Mixture-of-experts, 35B total, 3B active (256 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 22.5 GB / Q8 38.5 GB / FP16 70.9 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Agnes-AI/Agnes-3.0-Qwen](https://huggingface.co/Agnes-AI/Agnes-3.0-Qwen) · 2026-10-10 · qwen-community-1.0
  agnes-3.0-qwen is a post-trained derivative of Qwen/Qwen3.8-Flash-Next. Mixture-of-experts, 180B total (512 experts, 10 per token). Context: 1M tokens (config). License: qwen-community-1.0. Vendor-reported: "Reliable instruction following: on IFBench, agnes-3.0-qwen outperforms GPT-5.6 Sol (Max) and Claude Opus 4.8 (Max)."
  Runs on: need Q4 128.9 GB / Q8 192.8 GB / FP16 361.0 GB; smallest fit Q4: Mac 192GB, Q8: 2× DGX Spark (256GB), FP16: Mac 512GB
- [DataScience-UIBK/uibk-embedding-v1](https://huggingface.co/DataScience-UIBK/uibk-embedding-v1) · 2026-10-09 · apache-2.0
  UIBK-Embedding does it inside one model: - one encoder pass produces the token vectors and the dense embedding, - the dense embedding is stored as three extra vectors in the same list as the token vectors, - a plain MaxSim over that list is the sum of the two scores. Dense, 149M params. Context: 8K tokens (config). License: apache-2.0. Vendor-reported: "Parameters: 179M (shared trunk and token top 149.0M, dense top 20.1M, output layer 9.9M)"
  Runs on: need Q4 1.7 GB / Q8 1.8 GB / FP16 1.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [togethercomputer/HybridAgent-8B-Agentic](https://huggingface.co/togethercomputer/HybridAgent-8B-Agentic) · 2026-10-09 · mit
  HybridAgent is an LLM agent that dynamically navigates between latent reasoning and text-based interaction within a single generation process. Dense, 8.2B params. Context: 32K tokens (config). License: mit.
  Runs on: need Q4 7.2 GB / Q8 10.9 GB / FP16 18.6 GB; smallest fit Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090)
- [togethercomputer/HybridAgent-14B-Agentic](https://huggingface.co/togethercomputer/HybridAgent-14B-Agentic) · 2026-10-09 · mit
  HybridAgent is an LLM agent that dynamically navigates between latent reasoning and text-based interaction within a single generation process. Dense, 15B params. Context: 32K tokens (config). License: mit.
  Runs on: need Q4 11.6 GB / Q8 18.0 GB / FP16 31.8 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [alphaedge-ai/sparseup-fr-v0](https://huggingface.co/alphaedge-ai/sparseup-fr-v0) · 2026-10-08 · apache-2.0
  Feature extraction model from alphaedge-ai (HF pipeline tag; the card has no one-line description). Dense, 150M params. Context: 8K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.7 GB / Q8 1.8 GB / FP16 1.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [lightonai/LightOnOCR-3-1B](https://huggingface.co/lightonai/LightOnOCR-3-1B) · 2026-10-07 · apache-2.0
  LightOnOCR-3 is a new family of highly performant lightweight OCR models. Dense, 1B params. Context: 16K tokens (config). License: apache-2.0.
  Runs on: need Q4 2.5 GB / Q8 3.0 GB / FP16 3.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ifradlin/arrow](https://huggingface.co/ifradlin/arrow) · 2026-10-07 · cc-by-nc-4.0
  ARROW is a feed-forward model that unifies 3D reconstruction and 3D point tracking from arbitrary image sets. 1.3B params. License: cc-by-nc-4.0.
  Runs on: need Q4 2.3 GB / Q8 2.9 GB / FP16 4.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [egomnia/emma-6](https://huggingface.co/egomnia/emma-6) · 2026-10-07 · openrail
  Emma-6 is an open-source Italian Large Language Model developed by Egomnia S.p.A. Dense. License: openrail.
- [nvidia/cosmos3-nano-libero10-sdg-export-S2](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S2) · 2026-10-07 · nvidia-open-model-license
  Robotics model from nvidia (HF pipeline tag; the card has no one-line description). 15B params. License: nvidia-open-model-license.
  Runs on: need Q4 11.4 GB / Q8 18.4 GB / FP16 33.8 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [KelvinAxhcar/BitCache-Qwen2.5-0.5B-Instruct](https://huggingface.co/KelvinAxhcar/BitCache-Qwen2.5-0.5B-Instruct) · 2026-10-07 · apache-2.0
  BitCache-Qwen2.5-0.5B-Instruct is an optimized deployment of Alibaba Cloud's Qwen2.5 architecture powered by BitCache Spectral Attention Dynamics. Dense, 494M params. Context: 32K tokens (config). License: apache-2.0. Vendor-reported: "BitCache-Qwen2.5-0.5B-Instruct is an optimized deployment of Alibaba Cloud's state-of-the-art Qwen2.5 architecture powered by BitCache Spectral Attention Dynamics."
  Runs on: need Q4 1.5 GB / Q8 1.6 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [nvidia/cosmos3-nano-libero10-sdg-export-D2](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-D2) · 2026-10-07 · nvidia-open-model-license
  Robotics model from nvidia (HF pipeline tag; the card has no one-line description). 15B params. License: nvidia-open-model-license.
  Runs on: need Q4 11.4 GB / Q8 18.4 GB / FP16 33.8 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [nvidia/cosmos3-nano-libero10-sdg-export-S_all](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S_all) · 2026-10-07 · nvidia-open-model-license
  Robotics model from nvidia (HF pipeline tag; the card has no one-line description). 15B params. License: nvidia-open-model-license.
  Runs on: need Q4 11.4 GB / Q8 18.4 GB / FP16 33.8 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [nvidia/cosmos3-nano-libero10-sdg-export-D_half](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-D_half) · 2026-10-07 · nvidia-open-model-license
  Robotics model from nvidia (HF pipeline tag; the card has no one-line description). 15B params. License: nvidia-open-model-license.
  Runs on: need Q4 11.4 GB / Q8 18.4 GB / FP16 33.8 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [bottlecapai/ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B](https://huggingface.co/bottlecapai/ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B) · 2026-10-07 · openmdw-1.1
  ThinkingCap uses 23% fewer thinking tokens and completes 25 tasks in that time, compared to 11 for the base model. Mixture-of-experts, 32B total, 3B active (128 experts, 6 per token). Context: 256K tokens (config). License: openmdw-1.1.
  Runs on: need Q4 26.8 GB / Q8 36.4 GB / FP16 67.3 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [nanoswe/nanoswe-192h-261006](https://huggingface.co/nanoswe/nanoswe-192h-261006) · 2026-10-06 · mit
  A 1.7B-parameter SWE agent model trained from random init in 192 B200-hours (one 8×B200 node, 24.0 h of training wall-clock) in two stages: a short web-text pretraining stage (6.7B tokens of FineWeb-Edu), then 37.8B tokens of multi-turn tool-calling SWE agent trajectories — no instruction tuning… Dense, 2.8B params. Context: 32K tokens (config). License: mit. Vendor-reported: "It resolves 25.47% of SWE-bench Verified (pass@1 over K=5 samples on the 483-instance cluster-runnable subset), the new record for the 192 B200-hour budget."
  Runs on: need Q4 4.8 GB / Q8 6.1 GB / FP16 8.7 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [nvidia/cosmos3-nano-libero10-sdg-export-D](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-D) · 2026-10-06 · nvidia-open-model-license
  Robotics model from nvidia (HF pipeline tag; the card has no one-line description). 15B params. License: nvidia-open-model-license.
  Runs on: need Q4 11.4 GB / Q8 18.4 GB / FP16 33.8 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [nvidia/cosmos3-nano-libero10-sdg-export-R](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-R) · 2026-10-06 · nvidia-open-model-license
  Robotics model from nvidia (HF pipeline tag; the card has no one-line description). 15B params. License: nvidia-open-model-license.
  Runs on: need Q4 11.4 GB / Q8 18.4 GB / FP16 33.8 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000)
- [neuphonic/neudecide](https://huggingface.co/neuphonic/neudecide) · 2026-10-06 · apache-2.0
  NeuDecide is a pre-trained streaming English speech encoder joined to a compact tool-calling encoder–decoder, initialised from Needle, and trained end to end as one system. License: apache-2.0. Vendor-reported: "On SLURP, NeuDecide beats every other system on both metrics."
- [nvidia/lidarverse](https://huggingface.co/nvidia/lidarverse) · 2026-10-06 · openmdw-1.1
  LidarVerse is a real-time unified LiDAR world model for generation, point-cloud forecasting, and generalizable resimulation. License: openmdw-1.1. Vendor-reported: "LidarVerse achieves state-of-the-art performance on most reported metrics across the three evaluation protocols."
- [drivendata/pasketti-phonetic](https://huggingface.co/drivendata/pasketti-phonetic) · 2026-10-06 · bigcode-openrail-m
  pasketti-phonetic is an automatic speech recognition (ASR) model designed specifically for child speech. pasketti-phonetic generates phonetic (per-sound) transcriptions in the International Phonetic Alphabet (IPA), including nasalization and diphthong markings, from audio files. Dense, 316M params. License: bigcode-openrail-m.
  Runs on: need Q4 2.0 GB / Q8 2.1 GB / FP16 2.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [openbmb/MiniCPM-V-4.7-1B](https://huggingface.co/openbmb/MiniCPM-V-4.7-1B) · 2026-10-06 · apache-2.0
  MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 1B variant. Dense, 1.6B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "In Instruct mode, the 1B variant continues to improve foundation capabilities over MiniCPM-V 4.6 and substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency." Vendor-reported: "In Instruct mode, 1B substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency."
  Runs on: need Q4 2.1 GB / Q8 2.5 GB / FP16 3.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ncannings/parakeet-ultra-nemo](https://huggingface.co/ncannings/parakeet-ultra-nemo) · 2026-10-05 · cc-by-4.0
  Automatic speech recognition model from ncannings (HF pipeline tag; the card has no one-line description). License: cc-by-4.0.
- [FastVideo/FastVideo-FastH3-Omni-8-Step-V1](https://huggingface.co/FastVideo/FastVideo-FastH3-Omni-8-Step-V1) · 2026-10-05 · minimax-h3-community-license-agreement
  FastH3 Omni was built in close collaboration with Nuva Lab, which co-developed the model with FastVideo, defined and curated the omni reference dataset, and brought production-grounded creative video workflows to FastH3. License: minimax-h3-community-license-agreement.
  Runs on: need Q4 31.7 GB; smallest fit Q4: 48GB GPU (RTX 6000 Ada / A6000)
- [Meridian-MRM/Zedev-134m](https://huggingface.co/Meridian-MRM/Zedev-134m) · 2026-10-05 · apache-2.0
  Text generation model from Meridian-MRM (HF pipeline tag; the card has no one-line description). Dense, 134M params. Context: 1K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.2 GB / Q8 1.2 GB / FP16 1.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [BlackfireAI/nepali-asr-xlsr-300m](https://huggingface.co/BlackfireAI/nepali-asr-xlsr-300m) · 2026-10-04 · cc-by-nc-sa-4.0
  Automatic speech recognition model from BlackfireAI (HF pipeline tag; the card has no one-line description). Dense, 316M params. License: cc-by-nc-sa-4.0.
  Runs on: need Q4 2.0 GB / Q8 2.1 GB / FP16 2.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [BlackfireAI/nepali-asr-aud8-325m](https://huggingface.co/BlackfireAI/nepali-asr-aud8-325m) · 2026-10-04 · cc-by-nc-4.0
  Automatic speech recognition model from BlackfireAI (HF pipeline tag; the card has no one-line description). Dense, 325M params. Context: 32K tokens (config). License: cc-by-nc-4.0.
  Runs on: need Q4 1.3 GB / Q8 1.4 GB / FP16 1.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [drinnird/parakeet-tdt-0.6b-v3-sm8850-v81-fp16](https://huggingface.co/drinnird/parakeet-tdt-0.6b-v3-sm8850-v81-fp16) · 2026-10-04 · cc-by-4.0
  Automatic speech recognition model from drinnird (HF pipeline tag; the card has no one-line description). License: cc-by-4.0.
- [Blockway/Agens-Volundr-32B-Preview](https://huggingface.co/Blockway/Agens-Volundr-32B-Preview) · 2026-10-03 · apache-2.0
  Agens Volundr 32B is an open-weights model from Blockway, built for reasoning, coding and agent work. Dense, 33B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 30.0 GB / Q8 39.7 GB / FP16 67.9 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [Blockway/Agens-Volundr-32B-Preview-DFlash2](https://huggingface.co/Blockway/Agens-Volundr-32B-Preview-DFlash2) · 2026-10-03 · apache-2.0
  Text generation model from Blockway (HF pipeline tag; the card has no one-line description). Dense, 1.9B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 2.4 GB / Q8 3.2 GB / FP16 5.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [interfaze-ai/interfaze-1-lite](https://huggingface.co/interfaze-ai/interfaze-1-lite) · 2026-10-03 · interfaze-open
  Interfaze 1.1 Lite is a mixture-of-architectures (MoA) model for developer workloads: reading documents, transcribing speech, locating objects and interface elements, and answering over all of it with structured output. License: interfaze-open.
- [AneekC/quipu-moe-1B-A149M](https://huggingface.co/AneekC/quipu-moe-1B-A149M) · 2026-10-02 · apache-2.0
  Text generation model from AneekC (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 998M total, 149M active (64 experts). License: apache-2.0.
  Runs on: need Q4 2.1 GB / Q8 2.6 GB / FP16 3.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [ai-forever/FRIDA-Decisions](https://huggingface.co/ai-forever/FRIDA-Decisions) · 2026-10-02 · mit
  MlxJudge (frida-decisions[mlx], contributed by @akolotov) runs the model in MLX, without torch. Dense, 823M params. License: mit.
  Runs on: need Q4 2.0 GB / Q8 2.4 GB / FP16 3.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Aleph-Alpha/Kolibri-1-BF16](https://huggingface.co/Aleph-Alpha/Kolibri-1-BF16) · 2026-10-02 · apache-2.0
  Kolibri is Aleph Alpha's mixture-of-experts (MoE) reasoning model, with a focus on German and English. Mixture-of-experts, 78B total (384 experts, 6 per token). Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 51.6 GB / Q8 84.9 GB / FP16 158.1 GB; smallest fit Q4: Mac 96GB, Q8: 96GB GPU (RTX Pro 6000 Blackwell), FP16: 2× DGX Spark (256GB)
- [SoMarkAI/OXR-1.0](https://huggingface.co/SoMarkAI/OXR-1.0) · 2026-10-02
  OXR (Optical Everything Recognition) is an open-source model from the SoMark team. Dense, 1.5B params. Context: 40K tokens (config). Vendor-reported: "An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training." Vendor-reported: "With just 1.3B parameters, OXR ranks #1 on olmOCR-Bench, OmniDocBench, and DocParsingBench while delivering industry-leading parsing speed."
  Runs on: need Q4 2.4 GB / Q8 3.1 GB / FP16 4.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [mertkayacs/Tholos-2B](https://huggingface.co/mertkayacs/Tholos-2B) · 2026-10-01 · apache-2.0
  Tholos-2B is a 2B language model that chooses tools for AI assistants working with tables, notes and tasks. Dense, 2.5B params. Context: 128K tokens (config). License: apache-2.0.
  Runs on: need Q4 3.0 GB / Q8 4.1 GB / FP16 6.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [nvidia/PixelDiT2-ImageNet](https://huggingface.co/nvidia/PixelDiT2-ImageNet) · 2026-10-01
- [madrisight/madrimed1.2-VL-2B](https://huggingface.co/madrisight/madrimed1.2-VL-2B) · 2026-10-01 · apache-2.0
  MadriMed 1.2 is the second compact multimodal medical vision-language model by Madrisight. Dense, 2.4B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "It Delivers robust clinical reasoning capabilities, drastically reducing hallucinations and diagnostic contradictions while setting a new state-of-the-art benchmark for 2B medical vision models."
  Runs on: need Q4 3.2 GB / Q8 4.1 GB / FP16 6.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [cantina-security/apex-flash-1](https://huggingface.co/cantina-security/apex-flash-1) · 2026-09-30 · mit
  apex-flash-1 is Cantina Security's first open-weights security model, developed in partnership with Yeta (@yetalabs on X). Mixture-of-experts, 321B total (288 experts, 8 per token). Context: 1M tokens (config). License: mit.
  Runs on: need Q4 194.2 GB / Q8 342.8 GB / FP16 644.1 GB; smallest fit Q4: 2× DGX Spark (256GB), Q8: Mac 512GB
- [TelperionAI/Qwen3.8-27B-MLA](https://huggingface.co/TelperionAI/Qwen3.8-27B-MLA) · 2026-09-30 · apache-2.0
  Qwen3.8-27B is a hybrid model: 48 Gated DeltaNet (linear attention) layers and 16 full-attention layers with grouped-query attention (24 query heads, 4 KV heads, head dim 256). Dense, 27B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "Fidelity to the bf16 base model, teacher-forced on 562k held-out positions: top-1 disagreement where the base model is confident (top-1 vs top-2 logprob margin 2-5; lower is closer)."
  Runs on: need Q4 19.3 GB / Q8 31.7 GB / FP16 58.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [tfwnotops/Ternary-Bonsai-2-27B-ConstantKV](https://huggingface.co/tfwnotops/Ternary-Bonsai-2-27B-ConstantKV) · 2026-09-30 · apache-2.0
  Text generation model from tfwnotops (HF pipeline tag; the card has no one-line description). Dense, 27B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 19.5 GB / Q8 32.2 GB / FP16 57.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [nvidia/cosmos3-nano-libero10-sft-exports](https://huggingface.co/nvidia/cosmos3-nano-libero10-sft-exports) · 2026-09-29 · nvidia-open-model-license
  Robotics model from nvidia (HF pipeline tag; the card has no one-line description). License: nvidia-open-model-license.
- [inception42/Jais-3-32B](https://huggingface.co/inception42/Jais-3-32B) · 2026-09-29 · apache-2.0
  Jais-3-32B is the large dense member of the Jais-3 family: a 32B-class decoder-only model with a 512K context window. Dense, 35B params. Context: 512K tokens (config). License: apache-2.0.
  Runs on: need Q4 24.2 GB / Q8 40.1 GB / FP16 72.7 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [lighteternal/biodecision-v2-4b](https://huggingface.co/lighteternal/biodecision-v2-4b) · 2026-09-28 · research-only
  A System-1 decision model for biomedicine, pharma and clinical trials. Dense, 4.7B params. Context: 256K tokens (config). License: research-only.
  Runs on: need Q4 4.9 GB / Q8 7.1 GB / FP16 11.4 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [Lalakai/Hunter-1-4B](https://huggingface.co/Lalakai/Hunter-1-4B) · 2026-09-28 · apache-2.0
  Hunter 4B is a 4-billion-parameter conversational reasoning model from Lalakai Labs, built for local inference on consumer hardware. 4B params. License: apache-2.0.
  Runs on: need Q4 3.9 GB / Q8 5.7 GB / FP16 9.5 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060)
- [IndexTeam/Index-Translate-2B](https://huggingface.co/IndexTeam/Index-Translate-2B) · 2026-09-28 · apache-2.0
  Index-Translate-2B is the compact 2B-parameter multilingual text model of the Index-Translate family, built on Qwen3.5. Dense, 2.3B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "Outperforms Hy-MT2-1.8B on every main translation metric reported for both models in the technical report." Vendor-reported: "RIVAL provides adaptive judge supervision."
  Runs on: need Q4 2.7 GB / Q8 3.8 GB / FP16 6.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [NaiveAI/Naive-N0.5-Flash](https://huggingface.co/NaiveAI/Naive-N0.5-Flash) · 2026-09-27 · mit
  Naive-N0.5-Flash is an open-weight 309B MoE model with 15.5B active parameters, built for coding and AI R&D. Mixture-of-experts, 16B active (256 experts, 8 per token). Context: 1M tokens (config). License: mit. Vendor-reported: "SWA uses a 128-token window, while DSA selects the top 2,048 tokens for backbone attention."
  Runs on: need Q8 2.9 GB; smallest fit Q8: 8GB GPU (4060 / 3070)
- [shhivv/taiga-s1](https://huggingface.co/shhivv/taiga-s1) · 2026-09-27 · mit
  Taiga-S1 is the fast "System 1" layer for a CAD agent. 1M params. License: mit.
  Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [CentIo/Cent1-1B](https://huggingface.co/CentIo/Cent1-1B) · 2026-09-27 · apache-2.0
  Cent1-1B is a compact language model purpose-built for financial tool calling and numeric reasoning over financial statements. Dense, 1.7B params. Context: 40K tokens (config). License: apache-2.0.
  Runs on: need Q4 3.0 GB / Q8 3.7 GB / FP16 5.3 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Linzhan/UniMate](https://huggingface.co/Linzhan/UniMate) · 2026-09-27 · cc-by-nc-4.0
  License: cc-by-nc-4.0.
- [plasmova/nova-v3](https://huggingface.co/plasmova/nova-v3) · 2026-09-26 · apache-2.0
  Nova v3 is a decoder-only causal language model with 148.8 million parameters. Dense, 149M params. Context: 2K tokens (config). License: apache-2.0.
  Runs on: need Q4 1.4 GB / Q8 1.5 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [lighteternal/biodecision-tev1-4b](https://huggingface.co/lighteternal/biodecision-tev1-4b) · 2026-09-25 · research-only
  Together AI's Tev1 recipe: the model is trained to emit only the letter of the correct option (loss on the letter and end-of-turn token), which makes the letter distribution the answer distribution. Dense, 4.7B params. Context: 256K tokens (config). License: research-only. Vendor-reported: "Clinical-trial outcome forecasting (CT Open Endpoint): 72.6 and 62.0 macro-F1, on par with or above GPT-5, o3-mini and Claude Opus 4.5."
  Runs on: need Q4 4.9 GB / Q8 7.1 GB / FP16 11.4 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [HuggingFriends/RoboJudge-9B](https://huggingface.co/HuggingFriends/RoboJudge-9B) · 2026-09-25 · apache-2.0
  RoboJudge-9B is a multimodal judge for generated embodied-manipulation videos. Dense, 9.4B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 7.7 GB / Q8 12.1 GB / FP16 20.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [Comfy-Org/Ming-Image](https://huggingface.co/Comfy-Org/Ming-Image) · 2026-09-24 · mit
  License: mit.
- [madebyollin/texture-fix-vae-for-qwen-image-2.1](https://huggingface.co/madebyollin/texture-fix-vae-for-qwen-image-2.1) · 2026-09-24 · qwen-research
  Texture-Fix-VAE-for-Qwen-Image-2.1 is the Qwen-Image-2.1 VAE, but finetuned to produce cleaner textures with no checkerboard artifacts (and no NaNs when running in fp16). 338M params. License: qwen-research.
  Runs on: need Q4 1.7 GB / Q8 1.9 GB / FP16 2.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [FrancisRing/Prism](https://huggingface.co/FrancisRing/Prism) · 2026-09-24 · mit
  Image to video model from FrancisRing (HF pipeline tag; the card has no one-line description). License: mit. Vendor-reported: "Experiments show that Prism achieves 2.5$\times$ training speedup compared to full attention, while surpassing it in generation quality."
- [inclusionAI/AI-Transparency](https://huggingface.co/inclusionAI/AI-Transparency) · 2026-09-24
- [feyospace/feyospace-v1.1-small](https://huggingface.co/feyospace/feyospace-v1.1-small) · 2026-09-23
  Feyospace-v1.1-small is the updated small checkpoint in the Feyospace family of open-weight cyber agents, developed by the Feyospace Team at Vera Praxis. Dense, 28B params. Context: 256K tokens (config).
  Runs on: need Q4 19.8 GB / Q8 32.1 GB / FP16 58.7 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [mispeech/midashenglm-spatial](https://huggingface.co/mispeech/midashenglm-spatial) · 2026-09-23 · apache-2.0
  MiDashengLM-Spatial, to our knowledge, is the first open-source end-to-end unified audio-language model that supports both general audio understanding and spatial awareness within a single architecture. Dense, 8.4B params. Context: 32K tokens (config). License: apache-2.0.
  Runs on: need Q4 6.5 GB / Q8 10.4 GB / FP16 18.3 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090)
- [AxiomicLabs/GPT-X3-150M](https://huggingface.co/AxiomicLabs/GPT-X3-150M) · 2026-09-23 · apache-2.0
  GPT-X3-150M is a base model for text completion. Dense, 149M params. Context: 8K tokens (config). License: apache-2.0. Vendor-reported: "It ranks #1 on the Open SLM Leaderboard and is the first model under 150M parameters to exceed an Intelligence Index of 30." Vendor-reported: "GPT-X3-150M outperforms Hugging Face's SmolLM2-135M on the Intelligence Index while using about 27× fewer training tokens (75B vs 2T)."
  Runs on: need Q4 1.3 GB / Q8 1.4 GB / FP16 1.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [swiss-ai/Apertus-v1.5-8B-base](https://huggingface.co/swiss-ai/Apertus-v1.5-8B-base) · 2026-09-23 · apache-2.0
  Apertus 1.5 is a family of 8B and 70B fully open, multilingual, and multimodal models from the Swiss AI Initiative, trained only on openly available data and released with open weights, open data, and open training documentation. 8.8B params. License: apache-2.0.
  Runs on: need Q4 7.1 GB / Q8 11.2 GB / FP16 19.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [swiss-ai/Apertus-v1.5-70B-base](https://huggingface.co/swiss-ai/Apertus-v1.5-70B-base) · 2026-09-23 · apache-2.0
  Apertus 1.5 is a family of 8B and 70B fully open, multilingual, and multimodal models from the Swiss AI Initiative, trained only on openly available data and released with open weights, open data, and open training documentation. 72B params. License: apache-2.0.
  Runs on: need Q4 50.3 GB / Q8 83.5 GB / FP16 151.4 GB; smallest fit Q4: Mac 96GB, Q8: 96GB GPU (RTX Pro 6000 Blackwell), FP16: 2× DGX Spark (256GB)
- [SupersonicLabs/Julia-1](https://huggingface.co/SupersonicLabs/Julia-1) · 2026-09-23 · apache-2.0
  Julia 1 is the first model in the Julia family and the first released test of our training system. 144M params. License: apache-2.0.
  Runs on: need Q4 1.6 GB / Q8 1.7 GB / FP16 1.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [fastino/GLiNER2.5-Decide](https://huggingface.co/fastino/GLiNER2.5-Decide) · 2026-09-23 · apache-2.0
  Text classification model from fastino (HF pipeline tag; the card has no one-line description). 486M params. License: apache-2.0.
  Runs on: need Q4 1.8 GB / Q8 2.0 GB / FP16 2.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [bottlecapai/ThinkingCap-Qwen3.8-27B](https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B) · 2026-09-23 · polyform-small-business-1.0.0
  Image text to text model from bottlecapai (HF pipeline tag; the card has no one-line description). 28B params. License: polyform-small-business-1.0.0.
  Runs on: need Q4 20.8 GB / Q8 32.4 GB / FP16 58.1 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [ibm-granite/granite-timeseries-ensemble-r1](https://huggingface.co/ibm-granite/granite-timeseries-ensemble-r1) · 2026-09-23 · apache-2.0
  Granite Timeseries Ensemble R1 is an extensible framework for combining quantile forecasts from multiple pre-trained time-series models. License: apache-2.0. Vendor-reported: "Both ensembles outperform all four individual members on aggregate MASE and average MASE Rank." Vendor-reported: "In the 20 October 2026 comparison, the ensemble approach ranks 3rd among eligible systems on MASE Rank and CRPS Rank."
- [open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21](https://huggingface.co/open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21) · 2026-09-22 · openmdw-1.1
  Text generation model from open-athena (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 67B total, 2B active (256 experts, 4 per token). Context: 256K tokens (config). License: openmdw-1.1.
  Runs on: need Q4 41.7 GB / Q8 72.8 GB / FP16 135.7 GB; smallest fit Q4: 48GB GPU (RTX 6000 Ada / A6000), Q8: 96GB GPU (RTX Pro 6000 Blackwell), FP16: 2× DGX Spark (256GB)
- [black-forest-labs/flux-3-action-base](https://huggingface.co/black-forest-labs/flux-3-action-base) · 2026-09-22 · flux-kommunity-license
  FLUX 3 Action is an open weights 7B world action model. License: flux-kommunity-license.
- [moondream/parakeet-ultra](https://huggingface.co/moondream/parakeet-ultra) · 2026-09-22 · cc-by-4.0
  Its sibling Parakeet Redux is the ternary version of the same architecture, 178 MB, built for CPUs and Apple silicon. 627M params. License: cc-by-4.0.
  Runs on: need Q4 2.0 GB / Q8 2.2 GB / FP16 2.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [alibaba-pai/MiniMax-H3-Fun-Controlnet-Union-2.0](https://huggingface.co/alibaba-pai/MiniMax-H3-Fun-Controlnet-Union-2.0) · 2026-09-22 · minimax-h3-community-license-agreement
  Text to video model from alibaba-pai (HF pipeline tag; the card has no one-line description). License: minimax-h3-community-license-agreement.
- [black-forest-labs/flux-3-action-so101](https://huggingface.co/black-forest-labs/flux-3-action-so101) · 2026-09-22 · flux-kommunity-license
  FLUX 3 Action is an open weights 7B world action model. 6.9B params. License: flux-kommunity-license.
  Runs on: need Q4 5.8 GB / Q8 9.0 GB / FP16 15.5 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090)
- [black-forest-labs/flux-3-action-droid](https://huggingface.co/black-forest-labs/flux-3-action-droid) · 2026-09-22 · flux-kommunity-license
  FLUX 3 Action is an open weights 7B world action model. 6.9B params. License: flux-kommunity-license.
  Runs on: need Q4 5.8 GB / Q8 9.0 GB / FP16 15.5 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090)
- [inclusionAI/Ming-3.0-tts-0.5B](https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B) · 2026-09-22 · mit
  Text to speech model from inclusionAI (HF pipeline tag; the card has no one-line description). Dense, 1.5B params. Context: 32K tokens (config). License: mit.
  Runs on: need Q4 2.4 GB / Q8 3.2 GB / FP16 4.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [inclusionAI/Ming-3.0-tts-tiny](https://huggingface.co/inclusionAI/Ming-3.0-tts-tiny) · 2026-09-22 · mit
  Text to speech model from inclusionAI (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 9B total, 1.3B active (128 experts, 8 per token). Context: 4K tokens (config). License: mit.
  Runs on: need Q4 8.5 GB / Q8 12.4 GB / FP16 19.8 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [Winnougan/Cobijada_Minimax-H3_Hybrid_Pruned_ComfyUI](https://huggingface.co/Winnougan/Cobijada_Minimax-H3_Hybrid_Pruned_ComfyUI) · 2026-09-22 · apache-2.0
  Image to video model from Winnougan (HF pipeline tag; the card has no one-line description). License: apache-2.0.
- [paradigma-inc/limite-1b-violetto](https://huggingface.co/paradigma-inc/limite-1b-violetto) · 2026-09-21 · apache-2.0
  Limite is a 1-billion parameter dense autoregressive transformer, trained from scratch on a mixture of less than 300B highly curated tokens, capable of handling sequences up to 131k tokens long. Dense, 1B params. Context: 128K tokens (config). License: apache-2.0. Vendor-reported: "We leverage a mix of synthetic data generation, curated SFT and RL post-training to achieve results on competition-level math that rivals recent models tens of times larger, trained with orders of magnitudes more FLOPs."
  Runs on: need Q4 2.1 GB / Q8 2.6 GB / FP16 3.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [Contrastive-LM/CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) · 2026-09-21 · apache-2.0
  Text ranking model from Contrastive-LM (HF pipeline tag; the card has no one-line description). 8B params. License: apache-2.0. Vendor-reported: "Zero-shot: on par with Jev on computer-use, gaming and tool-calling tasks, with up to 9× lower latency." Vendor-reported: "Fine-tuned as a verifier: SOTA on DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%), 4–6× faster than Jev."
  Runs on: need Q4 6.5 GB / Q8 10.2 GB / FP16 18.1 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090)
- [XiaomiMiMo/MiMo-V2.6-Pro-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) · 2026-09-21 · mit
  MiMo-V2.6-Pro-RL is the flagship checkpoint of the MiMo-V2.6 series. Mixture-of-experts, 1024B total, 42B active (384 experts, 8 per token). Context: 1M tokens (config). License: mit.
  Runs on: need Q4 619.0 GB / Q8 1092.7 GB / FP16 2039.3 GB
- [XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B) · 2026-09-21 · mit
  MiMo-V2.6-Distill-Qwen-9B is a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data. Dense, 9.4B params. Context: 256K tokens (config). License: mit.
  Runs on: need Q4 7.9 GB / Q8 11.6 GB / FP16 20.0 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [XiaomiMiMo/MiMo-V2.6-Flash-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) · 2026-09-21 · mit
  MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of the MiMo-V2.6 series. Mixture-of-experts, 311B total, 15B active (256 experts, 8 per token). Context: 1M tokens (config). License: mit.
  Runs on: need Q4 188.7 GB / Q8 332.4 GB / FP16 623.7 GB; smallest fit Q4: 2× DGX Spark (256GB), Q8: Mac 512GB
- [apple/LensVLM-9B](https://huggingface.co/apple/LensVLM-9B) · 2026-09-21 · apple-amlr
  LensVLM is a 9B Vision Language Model (VLM) that scans compressed images of text, then selectively expands only the relevant pages to their uncompressed form via learned tools. Dense, 9.4B params. Context: 256K tokens (config). License: apple-amlr.
  Runs on: need Q4 7.9 GB / Q8 11.6 GB / FP16 20.0 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [microsoft/rho-roboeval](https://huggingface.co/microsoft/rho-roboeval) · 2026-09-21 · mit
  5.2B params. License: mit.
  Runs on: need Q4 4.6 GB / Q8 7.1 GB / FP16 12.0 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)
- [Altworld/Hemmingway-1](https://huggingface.co/Altworld/Hemmingway-1) · 2026-09-20 · cc-by-nc-4.0
  Text generation model from Altworld (HF pipeline tag; the card has no one-line description). Dense, 27B params. Context: 256K tokens (config). License: cc-by-nc-4.0. Vendor-reported: "It beats Fable 5.1, and it beats GPT-6 Astra by fifty points."
  Runs on: need Q4 20.5 GB / Q8 32.2 GB / FP16 57.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [JetBrains/Mellum2.1-12B-A2.5B-Thinking](https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking) · 2026-09-20 · apache-2.0
  Mellum2.1 is the next version of Mellum2 Thinking. Mixture-of-experts, 12B total, 2.5B active (64 experts, 8 per token). Context: 128K tokens (config). License: apache-2.0.
  Runs on: need Q4 9.6 GB / Q8 14.4 GB / FP16 25.8 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 32GB GPU (5090)
- [stepfun-ai/Step-5-Preview](https://huggingface.co/stepfun-ai/Step-5-Preview) · 2026-09-20 · other
  Step 5 Preview is a sparse Mixture-of-Experts (MoE) model with 600B total parameters and 27B active parameters. Mixture-of-experts, 604B total, 27B active (352 experts, 8 per token). Context: 1M tokens (config). License: other.
  Runs on: need Q4 400.6 GB / Q8 680.1 GB / FP16 1246.7 GB
- [Qwen/Qwen-Image-2.1-PE-T2I](https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I) · 2026-09-20 · qwen-research
  Qwen-Image-2.1 is a unified text-to-image generation and image editing model in the Qwen family. Dense, 9.4B params. Context: 256K tokens (config). License: qwen-research.
  Runs on: need Q4 7.7 GB / Q8 12.1 GB / FP16 20.0 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [Qwen/Qwen-Image-2.1-PE-I2I](https://huggingface.co/Qwen/Qwen-Image-2.1-PE-I2I) · 2026-09-20 · qwen-research
  Qwen-Image-2.1 is a unified text-to-image generation and image editing model in the Qwen family. Dense, 9.4B params. Context: 256K tokens (config). License: qwen-research.
  Runs on: need Q4 7.7 GB / Q8 12.1 GB / FP16 20.9 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090)
- [stepfun-ai/Step-5-Preview-BF16](https://huggingface.co/stepfun-ai/Step-5-Preview-BF16) · 2026-09-20 · other
  Step 5 Preview is a sparse Mixture-of-Experts (MoE) model with 600B total parameters and 27B active parameters. Mixture-of-experts, 604B total, 27B active (352 experts, 8 per token). Context: 1M tokens (config). License: other.
  Runs on: need Q4 400.6 GB / Q8 680.1 GB / FP16 1252.9 GB
- [moworld1/moworld_base](https://huggingface.co/moworld1/moworld_base) · 2026-09-20 · cc-by-nc-4.0
  MoWorld is a cost-efficient, high-performance Flash World Model for real-time interactive video generation on Neural Processing Units (NPUs). License: cc-by-nc-4.0.
- [HemmingwayAI/Hemmingway-1](https://huggingface.co/HemmingwayAI/Hemmingway-1) · 2026-09-20 · cc-by-nc-4.0
  Text generation model from HemmingwayAI (HF pipeline tag; the card has no one-line description). Dense, 27B params. Context: 256K tokens (config). License: cc-by-nc-4.0. Vendor-reported: "It beats Fable 5.1, and it beats GPT-6 Astra by fifty points."
  Runs on: need Q4 20.5 GB / Q8 32.2 GB / FP16 57.8 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [akhilaaa3/Jev-Omni](https://huggingface.co/akhilaaa3/Jev-Omni) · 2026-09-20 · apache-2.0
  Text classification model from akhilaaa3 (HF pipeline tag; the card has no one-line description). Dense, 12B params. Context: 256K tokens (config). License: apache-2.0.
  Runs on: need Q4 11.6 GB / Q8 16.9 GB / FP16 28.1 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 32GB GPU (5090)
- [openjev/openjev](https://huggingface.co/openjev/openjev) · 2026-09-20 · cc-by-nc-4.0
  OpenJev is an open-weights decision model. Dense, 27B params. Context: 256K tokens (config). License: cc-by-nc-4.0.
  Runs on: need Q4 19.6 GB / Q8 31.7 GB / FP16 56.9 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell)
- [convaiinnovations/laya-multilingual](https://huggingface.co/convaiinnovations/laya-multilingual) · 2026-09-19 · apache-2.0
  Text classification model from convaiinnovations (HF pipeline tag; the card has no one-line description). 322M params. License: apache-2.0.
  Runs on: need Q4 1.7 GB / Q8 1.9 GB / FP16 2.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya) · 2026-09-18 · apache-2.0
  Text classification model from convaiinnovations (HF pipeline tag; the card has no one-line description). 421M params. License: apache-2.0.
  Runs on: need Q4 1.8 GB / Q8 1.9 GB / FP16 2.3 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [moondream/parakeet-redux](https://huggingface.co/moondream/parakeet-redux) · 2026-09-18 · cc-by-4.0
  Its sibling Parakeet Ultra is the full-precision version of the same architecture, trained further, for GPUs: better than the original on every benchmark. 149M params. License: cc-by-4.0.
  Runs on: need Q4 1.6 GB / Q8 1.7 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [LiquidAI/LFM2.5-VL-3B-DSpark](https://huggingface.co/LiquidAI/LFM2.5-VL-3B-DSpark) · 2026-09-18 · lfm1.0
  LFM2.5-VL-3B-DSpark is an experimental speculative-decoding draft model that brings DSpark to vision-language models. Dense, 279M params. Context: 125K tokens (config). License: lfm1.0.
  Runs on: need Q4 1.3 GB / Q8 1.4 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [amazon/poiss-bge-m3-reranker](https://huggingface.co/amazon/poiss-bge-m3-reranker) · 2026-09-18 · apache-2.0
  Text classification model from amazon (HF pipeline tag; the card has no one-line description). Dense, 568M params. Context: 8K tokens (config). License: apache-2.0. Vendor-reported: "Negatives are the candidates the LLM judge marked non-relevant, padded from the reranked top-300 starting at rank 30."
  Runs on: need Q4 2.1 GB / Q8 2.4 GB / FP16 2.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [amazon/poiss-bge-m3-retriever](https://huggingface.co/amazon/poiss-bge-m3-retriever) · 2026-09-18 · mit
  Sentence similarity model from amazon (HF pipeline tag; the card has no one-line description). Dense, 567M params. Context: 8K tokens (config). License: mit. Vendor-reported: "Used as a reranker over its own top-1000 (the setting the cross-encoder is compared against): P@5 57.6, P@20 35.5, MRR 85.5, nDCG@5 61.5, nDCG@20 62.6."
  Runs on: need Q4 2.1 GB / Q8 2.4 GB / FP16 2.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070)
- [inclusionAI/Ming-Image-0.1-Design-Layer](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer) · 2026-09-17 · mit
  Ming-Image-0.1-Design-Layer decomposes a flattened design image into a requested number of RGBA layers using an image and a layer plan.
  Runs on: need Q4 5.2 GB / Q8 8.0 GB / FP16 13.8 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080)