# Local AI Release Board > Tracks new open-weight model releases (Hugging Face) and local-AI runtime/harness releases (GitHub), with a "runs on" fit check for NVIDIA/AMD/Intel Arc GPUs, DGX Spark, Strix Halo, Macs and CPU-only machines. Window: last 30 days; older items move to the archive. Generated 2026-10-24T23:23:29Z. - [Board](https://recentmodel.com/): Just landed + main board, filterable (devices, custom memory, GPU count, lowest quant, task, architecture, params, license, source); filters persist in the URL query string - [Archive](https://recentmodel.com/archive/): searchable archive of models that left the board (never deleted), same filters - [JSON feed](https://recentmodel.com/api/releases.json): models[], filtered_models[], harnesses[], ollama_newest[], x_posts[], devices[], estimate_note - [Archive JSON](https://recentmodel.com/api/archive.json): models[] with archive_reason, archived_at, low_signal, last-known stats and fit ## Notes - Facts come from the HF API/model cards, GitHub releases, ollama.com/library and public X posts. Unknown fields are left blank. - Main board threshold (non-trusted orgs must clear one): >= 50 HF likes, >= 5000 HF downloads, or >= 100 likes on X posts naming the model. Trusted orgs bypass it. Anything created or first seen in the last 72h skips the threshold and is shown as 'Just landed' (models[].lane=='fresh', vetted=false means 'new, unvetted'), sorted by velocity; unknown uploaders still need >= 2 likes or an X post linking the repo, and tests/checkpoints/merges/abliterated/re-uploads/empty cards/upload spam are dropped. After that window, items below threshold go to the archive marked low_signal (also listed in filtered_models[] for that run). - Items older than 30 days (release or first-seen) leave the board for the archive. - Advisory · OpenCode: Security incident Oct 17, 2026: rotate provider keys saved in OpenCode Console (https://ctxsec.com/items/opencode-console-provider-key-breach-2026-10) - Summaries are assembled from config.json and the model card; quoted claims are vendor-reported, not verified. - "Runs on": Need = weight size + FP16 KV cache for 8192 tokens (from config.json when available, otherwise a rough 8% of Q8 weights) + 1.0GB overhead. Weight size comes from the real quant files on Hugging Face when they exist (Q4_K_M, Q8_0 and BF16/F16 GGUF from the official org, unsloth, bartowski, ggml-org, lmstudio-community or mradermacher; MLX 4-bit/8-bit/bf16; or the official BF16 safetensors; another community GGUF only as a last resort, marked community), checked against the parameter count. When no matching file exists it is ESTIMATED at ~4.8/8.5/16 bits per weight and labelled 'est'. Q4 may use Unsloth UD-Q4_K_XL or Q4_K_S when no Q4_K_M exists. Usable memory per device is assumed below physical (GPUs ≈ VRAM − 1GB; Mac ≈75% of unified RAM; Strix Halo ≈96GB; CPU = RAM minus OS, and slow). est.fit in the JSON is for one card per device; the page can multiply discrete GPUs by 2 or 4, which assumes the runtime splits the model across cards (tensor/pipeline parallel or llama.cpp layer split). Arc classes only count Q8/FP16, matching what performs well on Arc. MoE models still need all weights resident; active params only affect speed. Fit is still a sizing calculation, not verified by running. ## Markdown pages - [Board markdown](https://recentmodel.com/index.md) - [Archive markdown](https://recentmodel.com/archive/index.md) # Local AI Release Board New open-weight models and local-AI runtime releases, with a hardware fit check for GPUs, DGX Spark, Strix Halo, Macs and CPU. Window: last 30 days. Data: [https://recentmodel.com/api/releases.json](https://recentmodel.com/api/releases.json). Fit is a sizing calculation, not verified by running. ## Just landed (54) - [kindlingai/glm-5.3-exlr8-k2-k3-k4](https://huggingface.co/kindlingai/glm-5.3-exlr8-k2-k3-k4) · 2026-10-08 · just landed · glm-5.3 EXLR8 is a new quantization format. Mixture-of-experts (256 experts, 8 per token). Context: 1M tokens (config). License: glm-5.3. Vendor-reported: "GLM-5.3's top-8 routing turns small numeric differences into different experts: two BF16 runs that differ only in their batching score mean 0.11 (median 0.004) against each other on 32,768 positions of the held-out text." - [MiniMaxAI/MiniMax-M3.1-Flash](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) · 2026-10-17 · just landed · minimax-community MiniMax-M3.1-Flash is a native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity. Mixture-of-experts, 220B total, 23B active (128 experts, 4 per token). Context: 1M tokens (config). License: minimax-community. Runs on: need Q4 134.3 GB / Q8 261.1 GB / FP16 443.0 GB; smallest fit Q4: 2× DGX Spark (256GB), Q8: Mac 512GB - [kindlingai/glm-5.3-flash-exlr8-k2-k3-k4](https://huggingface.co/kindlingai/glm-5.3-flash-exlr8-k2-k3-k4) · 2026-10-14 · just landed · mit GLM-5.3-Flash is a MoE model: 45 layers (0-2 dense, 3-44 MoE), 288 routed experts with top-8 routing plus one shared expert per MoE layer, an MTP layer 45 and a vision tower. Mixture-of-experts (288 experts, 8 per token). Context: 1M tokens (config). License: mit. Vendor-reported: "GLM-5.3-Flash is a MoE model: 45 layers (0-2 dense, 3-44 MoE), 288 routed experts with top-8 routing plus one shared expert per MoE layer, an MTP layer 45 and a vision tower." - [LiconStudio/Minimax-H3-VBVR](https://huggingface.co/LiconStudio/Minimax-H3-VBVR) · 2026-10-23 · just landed · apache-2.0 Licon VBVR is an image-to-video LoRA for MiniMax-H3, trained on VBVR data rebuilt and reprocessed specifically for H3. License: apache-2.0. - [tencent/WeVisOne-VL-4B](https://huggingface.co/tencent/WeVisOne-VL-4B) · 2026-10-22 · just landed · apache-2.0 WeVisOne-VL (WeChat Vision All-in-One) is a unified vision-language model developed by WeChat Vision, Tencent Inc. Dense, 4.8B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 4.9 GB / Q8 7.4 GB / FP16 11.9 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [HuggingFaceBio/Carbon-X](https://huggingface.co/HuggingFaceBio/Carbon-X) · 2026-10-21 · just landed · apache-2.0 Carbon-X is a 14.8B-parameter decoder-only language model trained jointly on DNA sequences and natural-language text. Dense, 15B params. Context: 40K tokens (config). License: apache-2.0. Runs on: need Q4 11.9 GB / Q8 18.1 GB / FP16 32.0 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000) - [RBDash-Team/Xing-VL1.0-Decision](https://huggingface.co/RBDash-Team/Xing-VL1.0-Decision) · 2026-10-22 · just landed, new/unvetted · apache-2.0 Xing-VL1.0-Decision is a 27B multimodal typed decision model. Dense, 26B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 19.6 GB / Q8 32.3 GB / FP16 55.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [tencent/WeVisOne-VL-2B](https://huggingface.co/tencent/WeVisOne-VL-2B) · 2026-10-22 · just landed · apache-2.0 WeVisOne-VL (WeChat Vision All-in-One) is a unified vision-language model developed by WeChat Vision, Tencent Inc. Dense, 2.4B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 3.2 GB / Q8 4.5 GB / FP16 6.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [sky7350/Mica-v0.5-4B](https://huggingface.co/sky7350/Mica-v0.5-4B) · 2026-09-27 · just landed, new/unvetted · apache-2.0 Mica is a 4B model that answers your questions about an input with a probability for every answer you allow. Dense, 4.7B params. Context: 256K tokens (config). License: apache-2.0. 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) - [bottlecapai/ThinkingCap-Qwen3.8-Flash-Next](https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-Flash-Next) · 2026-10-22 · just landed, new/unvetted · qwen-community-license-1.0 Thinking is on at reasoningeffort=xhigh, the chat template's own default, with the base model's recommended sampling — temperature=1.0, topp=0.95, topk=20, minp=0.0 — used unchanged for Ours. Mixture-of-experts, 180B total (512 experts, 10 per token). Context: 256K tokens (config). License: qwen-community-license-1.0. Runs on: need Q4 115.0 GB / Q8 192.8 GB / FP16 361.0 GB; smallest fit Q4: 1× DGX Spark (128GB), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB - [MinimaLabs/KLM-1.5](https://huggingface.co/MinimaLabs/KLM-1.5) · 2026-10-22 · just landed, new/unvetted · apache-2.0 KLM 1.5 is a 99M-parameter base language model. Dense, 99M params. Context: 2K tokens (config). License: apache-2.0. Runs on: need Q4 1.3 GB / Q8 1.3 GB / FP16 1.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16) · 2026-10-05 · just landed · openmdw-1.1 NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 is a small language model (SLM) further post-trained from Gemma-4-E4B-it by NVIDIA. Dense, 8B params. Context: 64K tokens (config). License: openmdw-1.1. Runs on: need Q4 6.8 GB / Q8 9.9 GB / FP16 16.8 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [inclusionAI/Ling-3.0-flash-Sante](https://huggingface.co/inclusionAI/Ling-3.0-flash-Sante) · 2026-10-19 · just landed · mit Text generation model from inclusionAI (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 127B total, 5.1B active (512 experts, 8 per token). Context: 256K tokens (config). License: mit. Runs on: need Q4 77.9 GB / Q8 137.8 GB / FP16 256.4 GB; smallest fit Q4: 96GB GPU (RTX Pro 6000 Blackwell), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB - [fastino/GLiDE](https://huggingface.co/fastino/GLiDE) · 2026-10-19 · just landed, new/unvetted · fastino-community GLiDE is a decision model from Fastino. Dense, 26B params. Context: 256K tokens (config). License: fastino-community. Runs on: need Q4 18.8 GB / Q8 30.8 GB / FP16 55.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [ltzheng/Qwen3.5-9B-General-Game-Stage1-3Shift-Shuffled](https://huggingface.co/ltzheng/Qwen3.5-9B-General-Game-Stage1-3Shift-Shuffled) · 2026-10-22 · just landed, new/unvetted · apache-2.0 Image text to text model from ltzheng (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 7.9 GB / Q8 12.4 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) - [darkc0de/XORTRON-AgentRICO](https://huggingface.co/darkc0de/XORTRON-AgentRICO) · 2026-10-19 · just landed, new/unvetted XORTRON is an experimental research system. Dense, 28B params. Context: 256K tokens (config). Runs on: need Q4 19.9 GB / Q8 32.8 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) - [som-doe2007/geoai-outdoor-assistant](https://huggingface.co/som-doe2007/geoai-outdoor-assistant) · 2026-10-23 · just landed, new/unvetted · apache-2.0 Text generation model from som-doe2007 (HF pipeline tag; the card has no one-line description). License: apache-2.0. - [bottlecapai/ThinkingCap-gemma-4-26B-A4B-it](https://huggingface.co/bottlecapai/ThinkingCap-gemma-4-26B-A4B-it) · 2026-10-23 · just landed, new/unvetted · apache-2.0 Image text to text model from bottlecapai (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 20.0 GB / Q8 29.9 GB / FP16 53.5 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [NorviTech/armiger-ultra](https://huggingface.co/NorviTech/armiger-ultra) · 2026-10-18 · just landed, new/unvetted · apache-2.0 Armiger Ultra is a 9-head byte-level RoPE Transformer (d=256, 6 layers, 8 heads) with a 6-plugin pipeline for raw BitTorrent/DHT metadata (name, file-tree paths + sizes, optional tracker/DHT counters). License: apache-2.0. - [Hikari07jp/DeepSeek-V4.1-1017-Ablitrated-testmodel](https://huggingface.co/Hikari07jp/DeepSeek-V4.1-1017-Ablitrated-testmodel) · 2026-10-23 · just landed, new/unvetted · mit Image text to text model from Hikari07jp (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 763B total (384 experts, 6 per token). Context: 1M tokens (config). License: mit. Vendor-reported: "Hidden size 5120, vocab 129,280, 384 routed experts (top-6) + 1 shared expert, MoE intermediate 2304." Runs on: need Q4 459.6 GB / Q8 812.6 GB / FP16 1528.1 GB - [tattrongvu/Jeffan-31b-1710](https://huggingface.co/tattrongvu/Jeffan-31b-1710) · 2026-10-19 · just landed, new/unvetted · apache-2.0 Text classification model from tattrongvu (HF pipeline tag; the card has no one-line description). Dense, 31B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "The Decision Index runs use the engine in inference/, which also reads a question a second time (the input shown twice) when its top-1 probability is below 0.9, and codes up to 255 options." 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) - [Valen-Team/Valen-0.8B](https://huggingface.co/Valen-Team/Valen-0.8B) · 2026-10-05 · just landed, new/unvetted · apache-2.0 Image text to text model from Valen-Team (HF pipeline tag; the card has no one-line description). Dense, 854M 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) - [nvidia/InternVideo2-CLIP-anomaly-detection](https://huggingface.co/nvidia/InternVideo2-CLIP-anomaly-detection) · 2026-10-22 · just landed · nvidia-internal-scientific-research-and-development-model-license Input Parameters: Video is a temporal sequence of two-dimensional RGB frames; text is a one-dimensional token sequence. 373M params. License: nvidia-internal-scientific-research-and-development-model-license. Vendor-reported: "The fine-tune improves the measured domain retrieval score (+8.78 percentage points mAP over the base) while reducing Kinetics-400 Top-1 accuracy by 0.92 percentage points versus the recorded base reference." Runs on: need Q4 1.7 GB / Q8 1.9 GB / FP16 2.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [ufakai/ufakzeka-kulak](https://huggingface.co/ufakai/ufakzeka-kulak) · 2026-10-22 · just landed, new/unvetted · apache-2.0 ufakzeka-kulak ("kulak" is Turkish for ear) is a small Turkish speech-to-text model by ufak AI, trained from scratch and released under Apache-2.0. License: apache-2.0. - [kakaocorp/kanana-2-vision-encoder-medium](https://huggingface.co/kakaocorp/kanana-2-vision-encoder-medium) · 2026-10-12 · just landed, new/unvetted · kanana-open-license Kanana-2-vision-encoder is Kakao's vision encoder series for vision-language models. Dense, 427M params. License: kanana-open-license. Runs on: need Q4 2.2 GB / Q8 2.4 GB / FP16 2.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [internlm/Atria-Dawn](https://huggingface.co/internlm/Atria-Dawn) · 2026-10-23 · just landed Atria Dawn is a new-generation agentic model developed by the Shanghai Artificial Intelligence Laboratory. Mixture-of-experts, 753B total (256 experts, 8 per token). Context: 1M tokens (config). Runs on: need Q4 453.7 GB / Q8 802.1 GB / FP16 1508.4 GB - [nvidia/GR00T-N1.7-AppleToPlate-cosine40k-checkpoint-15000](https://huggingface.co/nvidia/GR00T-N1.7-AppleToPlate-cosine40k-checkpoint-15000) · 2026-10-13 · just landed Robotics model from nvidia (HF pipeline tag; the card has no one-line description). 3.1B params. Runs on: need Q4 3.4 GB / Q8 4.8 GB / FP16 7.8 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060) - [ccortezb/FinOptix-14B-v2](https://huggingface.co/ccortezb/FinOptix-14B-v2) · 2026-10-23 · just landed, new/unvetted · apache-2.0 FinOptix-14B v2 is an open-weights FinOps specialist fine-tuned on Qwen 2.5 14B Instruct. Dense, 15B params. Context: 32K tokens (config). License: apache-2.0. Runs on: need Q4 11.6 GB / Q8 18.3 GB / FP16 32.1 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000) - [GrootTeam/Groot-19B-A3B-261023](https://huggingface.co/GrootTeam/Groot-19B-A3B-261023) · 2026-10-23 · just landed, new/unvetted · apache-2.0 Groot is a family of open-weight reasoning models for complex real-world tasks. Mixture-of-experts, 19B total, 3B active (128 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "Groot-19B-A3B-261023 beats both Qwen3.6-35B-A3B and Gemma-4-26B-A4B on 12 of the 16 benchmarks, for example Terminal Bench 2.1 (52.10 vs 43.82 / 26.52), NL2Repo-Bench (35.65 vs 28.77 / 16.55) and LongBench v2 (62.03 vs 48.12 / 47.99)." Runs on: need Q4 13.4 GB / Q8 22.5 GB / FP16 40.6 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), Q8: 24GB GPU (3090 / 4090), FP16: 48GB GPU (RTX 6000 Ada / A6000) - [SubMaroon/Kitchoon-V3-26B-A4B](https://huggingface.co/SubMaroon/Kitchoon-V3-26B-A4B) · 2026-10-20 · just landed, new/unvetted · apache-2.0 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: apache-2.0. 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) - [bayliner1980/Qwen3-0.6B-MoE-A0.5B](https://huggingface.co/bayliner1980/Qwen3-0.6B-MoE-A0.5B) · 2026-10-23 · just landed, new/unvetted · apache-2.0 Text generation model from bayliner1980 (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 597M total, 500M active (32 experts, 16 per token). Context: 40K tokens (config). License: apache-2.0. Vendor-reported: "At 0.6B, attention and the 156M-parameter embedding / LM head dominate the work per token, and top-16 activates half of the MLP." Runs on: need Q4 2.3 GB / Q8 2.5 GB / FP16 3.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [incoai/MiniMax-M3.1-Flash-DFlash2](https://huggingface.co/incoai/MiniMax-M3.1-Flash-DFlash2) · 2026-10-19 · just landed, new/unvetted · apache-2.0 DFlash 2 is a block-diffusion drafter for speculative decoding. Dense, 2.1B params. Context: 1M tokens (config). License: apache-2.0. Runs on: need Q4 2.5 GB / Q8 3.5 GB / FP16 5.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [ReadyArt/Wicked-Stuff-26B-A4B-v2.0-Final](https://huggingface.co/ReadyArt/Wicked-Stuff-26B-A4B-v2.0-Final) · 2026-10-23 · just landed, new/unvetted · 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 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) - [berryber09/FastH3-OmniRef-8step-pruned-w4a8](https://huggingface.co/berryber09/FastH3-OmniRef-8step-pruned-w4a8) · 2026-10-23 · just landed, new/unvetted · minimax-h3-community-license Text to video model from berryber09 (HF pipeline tag; the card has no one-line description). License: minimax-h3-community-license. - [GreenBitAI/Ternary-Bonsai-2-27B-gbx](https://huggingface.co/GreenBitAI/Ternary-Bonsai-2-27B-gbx) · 2026-10-20 · just landed, new/unvetted · apache-2.0 Bonsai 2 ships without a draft model; this head starts from the multi-token prediction head of Qwen3.8-27B, which Bonsai 2 is built from, and is distilled on text generated by Bonsai 2 itself -- prose, code, chat and coding-agent sessions. 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) - [Phase-Technologies/Densy-330M-Chat](https://huggingface.co/Phase-Technologies/Densy-330M-Chat) · 2026-10-23 · just landed, new/unvetted · apache-2.0 Densy-330M-Chat is an open-source, pure dense causal transformer language model designed and trained from scratch by Phase Technologies. 487M 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) - [owensong/Inflect-Mini-v3-Preview](https://huggingface.co/owensong/Inflect-Mini-v3-Preview) · 2026-10-23 · just landed, new/unvetted Text to speech model from owensong (HF pipeline tag; the card has no one-line description). - [tencent/Hy-Nomas](https://huggingface.co/tencent/Hy-Nomas) · 2026-10-21 · just landed · apache-2.0 Hy-Nomas is a video world generation model that supports free-form, multi-turn interaction through text instructions and camera control. License: apache-2.0. - [DavidAU/Qwen3.8-31B-Looping-Transformer](https://huggingface.co/DavidAU/Qwen3.8-31B-Looping-Transformer) · 2026-10-23 · just landed, new/unvetted · apache-2.0 Image text to text model from DavidAU (HF pipeline tag; the card has no one-line description). 31B params. License: apache-2.0. Runs on: need Q4 22.1 GB / Q8 36.4 GB / FP16 65.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [ReadyArt/Wicked-Stuff-31B-Final](https://huggingface.co/ReadyArt/Wicked-Stuff-31B-Final) · 2026-10-23 · just landed, new/unvetted · 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) - [mingqiancha233/Socrates-Image-1.0-31B](https://huggingface.co/mingqiancha233/Socrates-Image-1.0-31B) · 2026-10-22 · just landed, new/unvetted · qwen-research Socrates-Image-1.0 is a self-contained dense text-to-image model with recurrent visual-semantic feedback. 31B params. License: qwen-research. Runs on: need Q4 22.2 GB / Q8 36.5 GB / FP16 65.6 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [tencent/Hy-Nomas-Diffusers](https://huggingface.co/tencent/Hy-Nomas-Diffusers) · 2026-10-24 · just landed · apache-2.0 Hy-Nomas is a video world generation model that supports free-form, multi-turn interaction through text instructions and camera control. 8.6B params. License: apache-2.0. Runs on: need Q4 6.8 GB / Q8 10.8 GB / FP16 18.8 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [BlueRoseLabs/Silencio](https://huggingface.co/BlueRoseLabs/Silencio) · 2026-10-23 · just landed, new/unvetted · apache-2.0 Silencio is built on two open models. 34M params. 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) - [ReadyArt/Omega-Arsenal-26B-A4B-Final](https://huggingface.co/ReadyArt/Omega-Arsenal-26B-A4B-Final) · 2026-10-22 · just landed, new/unvetted · 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 20.7 GB / Q8 30.6 GB / FP16 54.6 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [h2oai/timeh2o-v1](https://huggingface.co/h2oai/timeh2o-v1) · 2026-10-22 · just landed, new/unvetted · h2o-model-non-commercial-licence-v1.0 Time series foundation model by H2O.ai. 109M params. License: h2o-model-non-commercial-licence-v1.0. Runs on: need Q4 1.6 GB / Q8 1.6 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [DavidAU/Qwen3.8-40B-Ultimate](https://huggingface.co/DavidAU/Qwen3.8-40B-Ultimate) · 2026-10-22 · just landed, new/unvetted · apache-2.0 Image text to text model from DavidAU (HF pipeline tag; the card has no one-line description). 40B params. License: apache-2.0. Runs on: need Q4 28.4 GB / Q8 46.9 GB / FP16 84.3 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [ATH-MaaS/PixText](https://huggingface.co/ATH-MaaS/PixText) · 2026-10-22 · just landed, new/unvetted · apache-2.0 Use models/PixText-Teacher to load Teacher. License: apache-2.0. - [ReadyArt/Wicked-Stuff-26B-A4B-v2.1-FINAL](https://huggingface.co/ReadyArt/Wicked-Stuff-26B-A4B-v2.1-FINAL) · 2026-10-24 · just landed, new/unvetted · 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 18.5 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) - [hi3d-official/Hi3Dv1.0](https://huggingface.co/hi3d-official/Hi3Dv1.0) · 2026-10-16 · just landed, new/unvetted · hi3d-open-license-agreement-1.0 Image to 3d model from hi3d-official (HF pipeline tag; the card has no one-line description). License: hi3d-open-license-agreement-1.0. - [e-n-v-y/None-qwenimage2.1-vae-comfyui](https://huggingface.co/e-n-v-y/None-qwenimage2.1-vae-comfyui) · 2026-10-21 · just landed, new/unvetted · qwen-research License: qwen-research. - [nikopueringer/CorridorKey2.0](https://huggingface.co/nikopueringer/CorridorKey2.0) · 2026-10-13 · just landed, new/unvetted · corridor-key-licence-2.0 CorridorKey runs on features from three frozen foundation models. License: corridor-key-licence-2.0. - [AnveshAI/AnveshAI-CoderLM-0.1B](https://huggingface.co/AnveshAI/AnveshAI-CoderLM-0.1B) · 2026-10-22 · just landed, new/unvetted · apache-2.0 CoderLM is a compact 100.2M-parameter language model purpose-built for multi-language code generation across 13 programming and markup languages — Python, Java, JavaScript, C++, C, Rust, TypeScript, Go, HTML, CSS, Shell, SQL, and Markdown. Dense, 100M params. Context: 2K tokens (config). License: apache-2.0. Runs on: need Q4 1.2 GB / Q8 1.2 GB / FP16 1.3 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [turion-ai/turion-jev-31b](https://huggingface.co/turion-ai/turion-jev-31b) · 2026-10-22 · just landed, new/unvetted · apache-2.0 Turion Jev 31B is a decision model by Turion AI, fine-tuned from google/gemma-4-31B-it on the Turion AI model training platform. 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) - [Luffuly/streaming-code-world-model](https://huggingface.co/Luffuly/streaming-code-world-model) · 2026-10-11 · just landed, new/unvetted ## Main board (131) - [RecursiveSI/flint](https://huggingface.co/RecursiveSI/flint) · 2026-10-21 · apache-2.0 Flint is Recursive’s mixture-of-experts language model for coding agents, with 39.8 billion total parameters and approximately 2.5 billion active per token. Mixture-of-experts, 40B total (256 experts, 6 per token). Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 26.9 GB / Q8 44.1 GB / FP16 81.4 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [KRAFTON/Raon-1-mini-preview](https://huggingface.co/KRAFTON/Raon-1-mini-preview) · 2026-10-21 · apache-2.0 Raon-1 mini is a 30B-A4B Mixture-of-Experts (MoE) language model, designed and trained from scratch by KRAFTON AI. Mixture-of-experts, 29B total (128 experts, 6 per token). Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 20.9 GB / Q8 32.6 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) - [inclusionAI/Ling-3.0-tiny-UBTree](https://huggingface.co/inclusionAI/Ling-3.0-tiny-UBTree) · 2026-10-21 · mit Text generation model from inclusionAI (HF pipeline tag; the card has no one-line description). Dense, 183M params. Context: 128K tokens (config). License: mit. Runs on: need Q4 1.4 GB / Q8 1.5 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [Comfy-Org/Fibo-2.0](https://huggingface.co/Comfy-Org/Fibo-2.0) · 2026-10-20 · cc-by-nc-4.0 License: cc-by-nc-4.0. - [FastVideo/FastH3-Omni-Comfy](https://huggingface.co/FastVideo/FastH3-Omni-Comfy) · 2026-10-20 · 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 14.7 GB; smallest fit Q4: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [XHToken/Spark-X2.6-4B](https://huggingface.co/XHToken/Spark-X2.6-4B) · 2026-10-19 · apache-2.0 Search for "Spark-X2.6" in LM Studio and download the model. 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 5.7 GB / Q8 6.6 GB / FP16 10.4 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060) - [lodestones/Kromoss](https://huggingface.co/lodestones/Kromoss) · 2026-10-19 · apache-2.0 Kromoss is a full fine-tune of MOSS-Image-Base at 1024 px. 5.1B params. License: apache-2.0. Runs on: need Q4 5.1 GB / Q8 7.2 GB / FP16 11.7 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [briaai/Fibo-2.0-edit-turbo](https://huggingface.co/briaai/Fibo-2.0-edit-turbo) · 2026-10-19 · bria-fibo Fibo 2.0 is the of Bria's FIBO family. 8.1B params. License: bria-fibo. Runs on: need Q4 6.6 GB / Q8 10.3 GB / FP16 18.0 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [gradium/phonon-31m-en](https://huggingface.co/gradium/phonon-31m-en) · 2026-10-19 · apache-2.0 phonon-31m is a text-to-speech model from the Phonon family. Dense, 41M params. Context: 8K tokens (config). License: apache-2.0. Runs on: need Q4 1.1 GB / Q8 1.1 GB / FP16 1.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [vllm-sr/d3-turbo](https://huggingface.co/vllm-sr/d3-turbo) · 2026-10-19 · apache-2.0 d3-turbo is the upgraded 27B multimodal decision model of Decision 3.0, the decision models of vLLM Semantic Router. Dense, 26B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 18.8 GB / Q8 30.8 GB / FP16 55.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [gradium/phonon-40m-en](https://huggingface.co/gradium/phonon-40m-en) · 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) - [nvidia/Cosmos3-H-Dreams](https://huggingface.co/nvidia/Cosmos3-H-Dreams) · 2026-10-19 · nvidia-open-model-license Cosmos3-H-Dreams is the successor of Cosmos-H-Dreams (arXiv:2608.24199), which was based on Cosmos-Predict2.5. License: nvidia-open-model-license. - [LiquidAI/LFM2.5-24B-A2B-DSpark](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-DSpark) · 2026-10-19 · lfm1.0 LFM2.5-24B-A2B-DSpark is a speculative-decoding draft model for LiquidAI/LFM2.5-24B-A2B. Dense, 328M params. Context: 128K tokens (config). License: lfm1.0. Runs on: need Q4 1.3 GB / Q8 1.4 GB / FP16 1.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/Cosmos3-H-Embodied-Nano](https://huggingface.co/nvidia/Cosmos3-H-Embodied-Nano) · 2026-10-19 · nvidia-open-model-license Cosmos3-H-Embodied-Nano is the successor of Cosmos-H-Surgical-Simulator, which was a forward-dynamics-only surgical simulator built on the 2B-parameter Cosmos-Predict2.5 diffusion transformer. License: nvidia-open-model-license. - [nvidia/cosmos3-nano-libero10-sdg-export-S2_s43](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S2_s43) · 2026-10-19 · 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-S5_2_s43](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S5_2_s43) · 2026-10-19 · 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-D2_s43](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-D2_s43) · 2026-10-19 · 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) - [briaai/Fibo-2.0-dev](https://huggingface.co/briaai/Fibo-2.0-dev) · 2026-10-18 · bria-fibo Fibo 2.0 is the of Bria's FIBO family. 8.1B params. License: bria-fibo. Runs on: need Q4 6.6 GB / Q8 10.3 GB / FP16 18.0 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [deepseek-ai/DeepSeek-V4.1-Flash-1017](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) · 2026-10-17 · mit DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks. Mixture-of-experts, 763B total (384 experts, 6 per token). Context: 1M tokens (config). License: mit. Runs on: need Q4 502.4 GB / Q8 510.1 GB / FP16 1528.1 GB - [autotrust/AutoTrust-JEV-27B-v1.1](https://huggingface.co/autotrust/AutoTrust-JEV-27B-v1.1) · 2026-10-17 · apache-2.0 AutoTrust-JEV-27B v1.1 is a text decision model that assigns probabilities to the answer options supplied with each question. 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) - [microsoft/Sentiment-Analysis-7B](https://huggingface.co/microsoft/Sentiment-Analysis-7B) · 2026-10-17 · mit Sentiment-Analysis-7B is a 7B-parameter autoregressive language model fine-tuned for sentiment classification of user messages directed at a service, product, or agent (the "Target"). Dense, 7.2B params. Context: 32K tokens (config). License: mit. Runs on: need Q4 6.5 GB / Q8 9.8 GB / FP16 16.6 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [Comfy-Org/MOSS-Image](https://huggingface.co/Comfy-Org/MOSS-Image) · 2026-10-16 · apache-2.0 License: apache-2.0. - [lodestones/MOSS-Image-SFT-8-steps-dmd](https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd) · 2026-10-16 · 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) - [nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16) · 2026-10-16 · openmdw-1.1 NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 is a looped (latent-recurrent) large language model (LLM) trained by NVIDIA. Dense, 4B params. Context: 256K tokens (config). License: openmdw-1.1. Runs on: need Q4 6.3 GB / Q8 7.6 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) - [ornith-ai/Ornith-1.5-122B](https://huggingface.co/ornith-ai/Ornith-1.5-122B) · 2026-10-16 · mit Text generation model from ornith-ai (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 125B total (256 experts, 8 per token). Context: 256K tokens (config). License: mit. Runs on: need Q4 77.6 GB / Q8 134.4 GB / FP16 251.2 GB; smallest fit Q4: 96GB GPU (RTX Pro 6000 Blackwell), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB - [tencent/Qwen3-8B-DSpine-b16](https://huggingface.co/tencent/Qwen3-8B-DSpine-b16) · 2026-10-16 · apache-2.0 DSpine is a parallel drafter for speculative decoding. Dense, 1.2B params. Context: 40K tokens (config). License: apache-2.0. Runs on: need Q4 1.9 GB / Q8 2.5 GB / FP16 3.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [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) - [microsoft/scFM-dataselection](https://huggingface.co/microsoft/scFM-dataselection) · 2026-10-16 · mit License: mit. - [leemysw/yovoice-model](https://huggingface.co/leemysw/yovoice-model) · 2026-10-16 · per-model 一般不需要手动下载：在 yovoice 的 设置 › 模型 中下载（或运行 yovoice models download ），应用会自动校验文件大小和 SHA256。. License: per-model. - [tencent/Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B) · 2026-10-15 · apache-2.0 Hy-Memory-30B-A3B is a compact long-term-memory writer model (~3B active parameters) developed by Tencent Hunyuan to execute the write path inside external memory frameworks. Mixture-of-experts, 30B total, 3B active (128 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 20.0 GB / Q8 33.8 GB / FP16 61.9 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [InternScience/Sage-A1.5](https://huggingface.co/InternScience/Sage-A1.5) · 2026-10-15 · apache-2.0 Sage‑A1.5 is a 35B Mixture‑of‑Experts agentic model based on Qwen3.6 and Agents-A1 from Shanghai AI Laboratory, built to advance long-horizon scientific problem solving through adaptive capability composition. Mixture-of-experts, 36B total (256 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 23.4 GB / Q8 39.5 GB / FP16 72.8 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [perplexity-ai/pplx-decider-v1.2](https://huggingface.co/perplexity-ai/pplx-decider-v1.2) · 2026-10-15 · apache-2.0 pplx-decider-v1.2 is the next version of our pplx-decider-v1.1-27b decision model, on the same Qwen3.8-27B backbone. Dense, 26B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 18.8 GB / Q8 30.8 GB / FP16 55.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [nvidia/Nemotron-3.5-Rerank-8B-BF16](https://huggingface.co/nvidia/Nemotron-3.5-Rerank-8B-BF16) · 2026-10-15 · openmdw-1.1 Text ranking model from nvidia (HF pipeline tag; the card has no one-line description). Dense, 8.4B params. Context: 256K tokens (config). License: openmdw-1.1. Vendor-reported: "Nemotron-3.5-Rerank-8B-BF16 achieves state-of-the-art results on the new ViDoRe v3.1 Reranking leaderboard) as of October 20, 2026." Runs on: need Q4 7.1 GB / Q8 11.0 GB / FP16 18.9 GB; smallest fit Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [perplexity-ai/pplx-decider-flash-v1.2](https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2) · 2026-10-15 · apache-2.0 pplx-decider-flash-v1.2 is the small version of pplx-decider-v1.2: the same joint request format, the same API and the same text and image support, at about one seventh of the size. Dense, 4.5B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 4.8 GB / Q8 7.3 GB / FP16 11.2 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [nvidia/RFD4-Proteina-prep](https://huggingface.co/nvidia/RFD4-Proteina-prep) · 2026-10-15 · bsd-3-clause License: bsd-3-clause. - [nvidia/RFD4-Proteina-Diffusers](https://huggingface.co/nvidia/RFD4-Proteina-Diffusers) · 2026-10-15 · bsd-3-clause Use the model for computational biomolecular design and follow designs with appropriate structural and experimental evaluation. 178M params. License: bsd-3-clause. Runs on: need Q4 1.6 GB / Q8 1.7 GB / FP16 1.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [black-forest-labs/flux-3-image-klein-comfy](https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy) · 2026-10-14 · flux-kommunity-license FLUX 3 Image Klein is a 5.6B open weights model for text-to-image generation and image editing. License: flux-kommunity-license. Runs on: need Q4 8.8 GB / Q8 14.1 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [mistralai/Timestral-209M](https://huggingface.co/mistralai/Timestral-209M) · 2026-10-14 · other Timestral is a family of time series foundation models for zero-shot forecasting, available in 5M, 28M, 123M, and 209M parameter versions. 209M params. License: other. Runs on: need Q4 1.6 GB / Q8 1.7 GB / FP16 1.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [mistralai/Timestral-5M](https://huggingface.co/mistralai/Timestral-5M) · 2026-10-14 · other Timestral is a family of time series foundation models for zero-shot forecasting, available in 5M, 28M, 123M, and 209M parameter versions. 5M params. License: other. Vendor-reported: "All Timestreal models outperform every other model at the respective model size." Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [black-forest-labs/flux-3-image-klein-pup](https://huggingface.co/black-forest-labs/flux-3-image-klein-pup) · 2026-10-14 · apache-2.0 FLUX 3 Image Klein Prompt Upsampler is a 9B vision-language model that expands a short image request into the detailed prompt FLUX 3 Image Klein works best with. Dense, 9.4B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 7.7 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) - [ibm-granite/granite-switch-4.1-8b](https://huggingface.co/ibm-granite/granite-switch-4.1-8b) · 2026-10-14 · apache-2.0 Granite Switch 4.1 8B is a modular LLM that composes a base model, multiple adapters and an encoder together so that a single hosted endpoint can move between natural-language responses, general and specialized decisions, and speech-to-text. Dense, 8.8B params. Context: 128K tokens (config). License: apache-2.0. Runs on: need Q4 8.8 GB / Q8 12.5 GB / FP16 20.4 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090) - [inclusionAI/Ming-3.0-asr-tiny](https://huggingface.co/inclusionAI/Ming-3.0-asr-tiny) · 2026-10-14 · mit Automatic speech recognition model from inclusionAI (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 8.9B total (128 experts, 8 per token). Context: 8K tokens (config). License: mit. Runs on: need Q4 6.5 GB / Q8 10.6 GB / FP16 18.9 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [mistralai/Timestral-123M](https://huggingface.co/mistralai/Timestral-123M) · 2026-10-14 · other Timestral is a family of time series foundation models for zero-shot forecasting, available in 5M, 28M, 123M, and 209M parameter versions. 123M params. License: other. Vendor-reported: "All Timestreal models outperform every other model at the respective model size." Runs on: need Q4 1.6 GB / Q8 1.6 GB / FP16 1.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [mistralai/Timestral-28M](https://huggingface.co/mistralai/Timestral-28M) · 2026-10-14 · other Timestral is a family of time series foundation models for zero-shot forecasting, available in 5M, 28M, 123M, and 209M parameter versions. 28M params. License: other. Vendor-reported: "All Timestreal models outperform every other model at the respective model size." Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [ibm-granite/granite-switch-4.1-30b](https://huggingface.co/ibm-granite/granite-switch-4.1-30b) · 2026-10-14 · apache-2.0 Granite Switch 4.1 30B is a modular LLM that composes a base model, multiple adapters and an encoder together so that a single hosted endpoint can move between natural-language responses, general and specialized decisions, and speech-to-text. Dense, 30B params. Context: 128K tokens (config). License: apache-2.0. Runs on: need Q4 24.2 GB / Q8 37.2 GB / FP16 64.3 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [microsoft/UserRepGen-4B](https://huggingface.co/microsoft/UserRepGen-4B) · 2026-10-14 · mit UserRepGen-4B is a fine-tuned Qwen3.5-4B language model that generates a structured, natural-language user representation from a multi-turn user-assistant conversation. Dense, 4.5B params. Context: 256K tokens (config). License: mit. Runs on: need Q4 4.8 GB / Q8 6.6 GB / FP16 10.5 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060) - [inclusionAI/Ling-3.0-flash-UBTree](https://huggingface.co/inclusionAI/Ling-3.0-flash-UBTree) · 2026-10-14 · mit Text generation model from inclusionAI (HF pipeline tag; the card has no one-line description). Dense, 482M params. Context: 128K tokens (config). License: mit. Runs on: need Q4 2.0 GB / Q8 2.2 GB / FP16 2.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [amazon/LAR-27B](https://huggingface.co/amazon/LAR-27B) · 2026-10-14 · apache-2.0 Dense, 27B params. Context: 256K tokens (config). License: apache-2.0. 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) - [amazon/LAR-9B](https://huggingface.co/amazon/LAR-9B) · 2026-10-14 · apache-2.0 Dense, 9B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 7.7 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) - [nvidia/cdf-sim2real-smt-solderlight](https://huggingface.co/nvidia/cdf-sim2real-smt-solderlight) · 2026-10-14 · openmdw-1.1 License: openmdw-1.1. - [black-forest-labs/flux-3-image-klein](https://huggingface.co/black-forest-labs/flux-3-image-klein) · 2026-10-13 · flux-kommunity-license FLUX 3 Image Klein is a 5.6B open weights model for text-to-image generation and image editing. 5.5B params. License: flux-kommunity-license. Runs on: need Q4 5.2 GB / Q8 7.6 GB / FP16 12.6 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [black-forest-labs/flux-3-image-klein-base](https://huggingface.co/black-forest-labs/flux-3-image-klein-base) · 2026-10-13 · flux-kommunity-license FLUX 3 Image Klein is a 5.6B open weights model for text-to-image generation and image editing. 5.5B params. License: flux-kommunity-license. Runs on: need Q4 5.1 GB / Q8 7.6 GB / FP16 12.6 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [inclusionAI/Ling-3.1-flash](https://huggingface.co/inclusionAI/Ling-3.1-flash) · 2026-10-13 · mit Text generation model from inclusionAI (HF pipeline tag; the card has no one-line description). Mixture-of-experts, 572B total (512 experts, 8 per token). Context: 256K tokens (config). License: mit. Runs on: need Q4 359.7 GB / Q8 610.1 GB / FP16 1145.9 GB; smallest fit Q4: Mac 512GB - [reflection/Beam-1.0-501B-A23B-DFlash2](https://huggingface.co/reflection/Beam-1.0-501B-A23B-DFlash2) · 2026-10-13 · apache-2.0 DFlash2 draft model for Beam. Dense, 1.9B params. Context: 1M tokens (config). License: apache-2.0. Runs on: need Q4 2.4 GB / Q8 3.2 GB / FP16 5.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/Kimodo-SOMA-RP-ArdyRep-v1.2](https://huggingface.co/nvidia/Kimodo-SOMA-RP-ArdyRep-v1.2) · 2026-10-13 · nvidia-open-model-license The Kimodo model family includes models trained on different skeletons and datasets: Kimodo-SOMA-RP Trained on the 30-joint SOMA skeleton with the proprietary Bones Rigplay dataset. 283M params. License: nvidia-open-model-license. Runs on: need Q4 1.7 GB / Q8 1.8 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/Kimodo-SOMA-RP-v1.2](https://huggingface.co/nvidia/Kimodo-SOMA-RP-v1.2) · 2026-10-13 · nvidia-open-model-license The Kimodo model family includes models trained on different skeletons and datasets: Kimodo-SOMA-RP Trained on the 30-joint SOMA skeleton with the proprietary Bones Rigplay dataset. 283M params. License: nvidia-open-model-license. Runs on: need Q4 1.7 GB / Q8 1.8 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/ARDY-SOMA-RP-30FPS-KimodoRep-Horizon60](https://huggingface.co/nvidia/ARDY-SOMA-RP-30FPS-KimodoRep-Horizon60) · 2026-10-13 · nvidia-open-model-agreement License: nvidia-open-model-agreement. - [nvidia/Cosmos-AnomalyGen-Glass-Nano](https://huggingface.co/nvidia/Cosmos-AnomalyGen-Glass-Nano) · 2026-10-13 · openmdw-1.1 License: openmdw-1.1. - [nvidia/Cosmos-AnomalyGen-Metal-Nano](https://huggingface.co/nvidia/Cosmos-AnomalyGen-Metal-Nano) · 2026-10-13 · openmdw-1.1 License: openmdw-1.1. - [nvidia/Cosmos-AnomalyGen-PCB-Nano](https://huggingface.co/nvidia/Cosmos-AnomalyGen-PCB-Nano) · 2026-10-13 · openmdw-1.1 License: openmdw-1.1. - [LiquidAI/LFM2.5-24B-A2B](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) · 2026-10-12 · lfm1.0 LFM2.5-24B-A2B is a 24B-parameter reasoning model with only 2B active parameters, built for fast, private agents on-device. Mixture-of-experts, 24B total, 2B active (64 experts, 4 per token). Context: 128K tokens (config). License: lfm1.0. Runs on: need Q4 16.2 GB / Q8 27.2 GB / FP16 49.7 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16) · 2026-10-12 · openmdw-1.1 NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 is the full-precision release of NVIDIA Nemotron 3.5 Super VL, a vision-language model for reasoning, coding, tool use, document understanding, and video understanding. Mixture-of-experts, 124B total, 12B active (512 experts, 22 per token). Context: 1M tokens (config). License: openmdw-1.1. Runs on: need Q4 98.5 GB / Q8 143.2 GB / FP16 261.2 GB; smallest fit Q4: 1× DGX Spark (128GB), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB - [nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16) · 2026-10-12 · openmdw-1.1 NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Base-BF16 is the base vision-language checkpoint of the Nemotron 3.5 Super family. Mixture-of-experts, 124B total, 12B active (512 experts, 22 per token). Context: 256K tokens (config). License: openmdw-1.1. Runs on: need Q4 86.3 GB / Q8 143.8 GB / FP16 260.5 GB; smallest fit Q4: 96GB GPU (RTX Pro 6000 Blackwell), Q8: 2× DGX Spark (256GB), FP16: Mac 512GB - [theforecastingcompany/t0-gamma](https://huggingface.co/theforecastingcompany/t0-gamma) · 2026-10-12 · apache-2.0 Time series forecasting model from theforecastingcompany (HF pipeline tag; the card has no one-line description). Dense, 256M params. License: apache-2.0. Runs on: need Q4 1.7 GB / Q8 1.8 GB / FP16 2.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [openbmb/Trident-Qwen3VL-2B](https://huggingface.co/openbmb/Trident-Qwen3VL-2B) · 2026-10-12 · mit Trident-Qwen3VL-2B is a multimodal embedding model for mixed-modality retrieval. Dense, 2.1B params. Context: 256K tokens (config). License: mit. Runs on: need Q4 3.2 GB / Q8 4.2 GB / FP16 6.2 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [allenai/HiRO-ACEv2](https://huggingface.co/allenai/HiRO-ACEv2) · 2026-10-12 · apache-2.0 HiRO (High Resolution Output) is a diffusion model which generates downscaled fields at 3 km resolution from 100 km resolution inputs. License: apache-2.0. - [OpenMOSS-Team/MOSS-Image-SFT](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) · 2026-10-11 · apache-2.0 MOSS-Image-SFT is a supervised fine-tuned text-to-image model in the MOSS-Image family from OpenMOSS, derived from MOSS-Image-Base. 5.1B params. License: apache-2.0. Runs on: need Q4 4.7 GB / Q8 7.0 GB / FP16 11.7 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [OpenMOSS-Team/MOSS-Image-Base](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base) · 2026-10-11 · apache-2.0 MOSS-Image-Base is the base text-to-image model in the MOSS-Image family from OpenMOSS. 5.1B params. License: apache-2.0. Runs on: need Q4 4.7 GB / Q8 7.0 GB / FP16 11.7 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [datalab-to/indra-ocr](https://huggingface.co/datalab-to/indra-ocr) · 2026-10-10 · openrail Indra OCR is a 1M-parameter OCR model that runs on CPU. 1M params. License: openrail. Vendor-reported: "When you need the highest accuracy, on handwriting, math, complex tables and degraded scans, our managed platform runs Chandra, our state of the art OCR model, with zero data retention by default, SOC 2 Type 2, and custom BAAs." Runs on: need Q4 1.5 GB / Q8 1.5 GB / FP16 1.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [vllm-sr/d3](https://huggingface.co/vllm-sr/d3) · 2026-10-10 · apache-2.0 Zero shot classification model from vllm-sr (HF pipeline tag; the card has no one-line description). Dense, 26B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "d3 ranks first of the 149 entries on the official text board and is on its Pareto frontier: no entry with as many parameters or fewer scores higher." Runs on: need Q4 22.5 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) - [dmis-lab/UltraBERT-380M](https://huggingface.co/dmis-lab/UltraBERT-380M) · 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, 430M 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 2.1 GB / Q8 2.3 GB / FP16 2.7 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [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) - [nvidia/cosmos3-nano-libero10-sdg-export-S5_2](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S5_2) · 2026-10-10 · 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-S5_all](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S5_all) · 2026-10-10 · 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) - [Qwen/Qwen-Image-2.1-Turbo](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo) · 2026-10-09 · qwen-research Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing with 8 denoising steps. 7.1B params. License: qwen-research. Runs on: need Q4 5.8 GB / Q8 9.2 GB / FP16 15.8 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [Cloudflare/clef-omni](https://huggingface.co/Cloudflare/clef-omni) · 2026-10-09 · apache-2.0 Clef-Omni is a 30B-A3B mixture-of-experts multimodal model that turns a state and a schema of typed questions into decisions. 35B params. License: apache-2.0. Runs on: need Q4 23.8 GB / Q8 39.0 GB / FP16 74.7 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [tencent/Youtu-Parsing-Omni](https://huggingface.co/tencent/Youtu-Parsing-Omni) · 2026-10-09 · youtu-parsing Youtu-Parsing-Omni is a compact (5B) omni-modal parsing model. Dense, 5.3B params. Context: 1M tokens (config). License: youtu-parsing. Runs on: need Q4 4.6 GB / Q8 7.1 GB / FP16 12.1 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16) · 2026-10-09 · openmdw-1.1 NVIDIA Nemotron 3.5 Bolt 74B-A9B is a vision-language model (VLM) designed to run everyday agentic tasks locally. Mixture-of-experts, 75B total, 9B active (512 experts). Context: 256K tokens (config). License: openmdw-1.1. Runs on: need Q4 57.6 GB / Q8 86.3 GB / FP16 155.6 GB; smallest fit Q4: Mac 96GB, Q8: 96GB GPU (RTX Pro 6000 Blackwell), FP16: 2× DGX Spark (256GB) - [LiquidAI/LFM2.5-24B-A2B-Base](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-Base) · 2026-10-09 · lfm1.0 LFM2.5 is a new family of hybrid models designed for on-device deployment. Mixture-of-experts, 24B total, 2B active (64 experts, 4 per token). Context: 128K tokens (config). License: lfm1.0. Runs on: need Q4 16.1 GB / Q8 27.2 GB / FP16 49.6 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [nvidia/cosmos3-nano-libero10-sdg-export-D5](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-D5) · 2026-10-09 · 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) - [cloud0day3/antalia-mini](https://huggingface.co/cloud0day3/antalia-mini) · 2026-10-08 · apache-2.0 Antalia-2 Mini (version 1.0.1) is a small Turkish text-to-speech model with one voice. License: apache-2.0. - [mistralai/Voxtral-Mini-4B-Realtime-Arabic](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-Arabic) · 2026-10-08 Voxtral Mini 4B Realtime Arabic is a streaming speech-to-text model for Arabic dialects and Modern Standard Arabic. Dense, 4.4B params. Context: 128K tokens (config). Vendor-reported: "It outperforms several open-source and proprietary offline models and comes within 0.91 percentage points of Voxtral Transcribe Arabic’s 7.91% CER." Runs on: need Q4 4.1 GB / Q8 6.6 GB / FP16 10.8 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060) - [mistralai/LIDstral-Arabic](https://huggingface.co/mistralai/LIDstral-Arabic) · 2026-10-08 · apache-2.0 LIDstral Arabic is a fast language and dialect classifier for text written in Arabic script. License: apache-2.0. - [openbmb/MiniCPM-V-4.7-35B-A3B-DSpark](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B-DSpark) · 2026-10-08 · apache-2.0 Text generation model from openbmb (HF pipeline tag; the card has no one-line description). Dense, 389M params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 1.3 GB / Q8 1.5 GB / FP16 1.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [openbmb/MiniCPM-V-4.7-1B-DSpark](https://huggingface.co/openbmb/MiniCPM-V-4.7-1B-DSpark) · 2026-10-08 · apache-2.0 Text generation model from openbmb (HF pipeline tag; the card has no one-line description). Dense, 214M params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 1.2 GB / Q8 1.3 GB / FP16 1.5 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/cosmos3-nano-libero10-sdg-export-S_best](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S_best) · 2026-10-08 · 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_worst](https://huggingface.co/nvidia/cosmos3-nano-libero10-sdg-export-S_worst) · 2026-10-08 · 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) - [openbmb/MiniCPM-V-4.7-35B-A3B](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B) · 2026-10-07 · apache-2.0 MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 35B-A3B variant, with 35B-A3B being the first MoE model in the MiniCPM-V series. Mixture-of-experts, 35B total, 3B active (256 experts, 8 per token). Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 22.9 GB / Q8 38.6 GB / FP16 71.9 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [reflection/Beam-1.0-501B-A23B](https://huggingface.co/reflection/Beam-1.0-501B-A23B) · 2026-10-07 · apache-2.0 Beam is an open model built by Reflection. Mixture-of-experts, 501B total, 23B active (256 experts, 8 per token). Context: 1M tokens (config). License: apache-2.0. Runs on: need Q4 309.2 GB / Q8 535.5 GB / FP16 1005.4 GB; smallest fit Q4: Mac 512GB - [lightonai/LightOnOCR-3-0.8B](https://huggingface.co/lightonai/LightOnOCR-3-0.8B) · 2026-10-07 · apache-2.0 LightOnOCR-3-0.8B is the lightest model of LightOn's LightOnOCR-3 family, with the same layout and visual understanding features. 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) - [ibm-granite/granite-speech-5.0-2b](https://huggingface.co/ibm-granite/granite-speech-5.0-2b) · 2026-10-07 · apache-2.0 Granite Speech 5.0 2B is a compact and efficient speech-language model, specifically designed for multilingual automatic speech recognition (ASR) and bidirectional automatic speech translation (AST) for English, French, German, Spanish, Portuguese and Japanese. Dense, 2.3B params. Context: 16K tokens (config). License: apache-2.0. Runs on: need Q4 2.8 GB / Q8 3.7 GB / FP16 5.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/Nemotron-3.5-Embed-1B-BF16](https://huggingface.co/nvidia/Nemotron-3.5-Embed-1B-BF16) · 2026-10-07 · openmdw-1.1 The Nemotron-3.5-Embed-1B-BF16 model was distilled from the Nemotron-3.5-Embed-8B-BF16 embedding teacher using combined cosine distance (COS) and mean squared error (MSE) losses on multilingual text and image retrieval data. Dense, 1.6B params. Context: 256K tokens (config). License: openmdw-1.1. Vendor-reported: "Among models of comparable size, Nemotron-3.5-Embed-1B-BF16 achieves state-of-the-art results across multiple multilingual and multimodal retrieval benchmarks." Runs on: need Q4 2.2 GB / Q8 2.7 GB / FP16 4.6 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [nvidia/Nemotron-3.5-Embed-8B-BF16](https://huggingface.co/nvidia/Nemotron-3.5-Embed-8B-BF16) · 2026-10-07 · openmdw-1.1 Visual document retrieval model from nvidia (HF pipeline tag; the card has no one-line description). Dense, 8.4B params. Context: 256K tokens (config). License: openmdw-1.1. Vendor-reported: "Nemotron-3.5-Embed-8B-BF16 achieves state-of-the-art results across multiple multilingual and multimodal retrieval benchmarks." Runs on: need Q4 7.0 GB / Q8 10.6 GB / FP16 18.9 GB; smallest fit Q4: 8GB GPU (4060 / 3070), Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [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) - [lightonai/LightOnOCR-3-4B](https://huggingface.co/lightonai/LightOnOCR-3-4B) · 2026-10-06 · apache-2.0 LightOnOCR-3-4B is the largest and most accurate model of the LightOnOCR-3 family, LightOn's new generation of lightweight end-to-end OCR models with layout and visual understanding. Dense, 4.5B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 4.8 GB / Q8 6.9 GB / FP16 11.2 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [nvidia/RFD4-Proteina](https://huggingface.co/nvidia/RFD4-Proteina) · 2026-10-06 · bsd-3-clause Use the RFD4 loader below; this is a native RFD4 bundle. 436M params. License: bsd-3-clause. Runs on: need Q4 1.8 GB / Q8 2.0 GB / FP16 2.4 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [LiquidAI/d1-3B](https://huggingface.co/LiquidAI/d1-3B) · 2026-10-05 · lfm1.0 d1-3B is a 3B parameter decision model built on LFM2.5-VL-3B. Dense, 3.1B params. Context: 32K tokens (config). License: lfm1.0. Runs on: need Q4 3.2 GB / Q8 4.4 GB / FP16 6.9 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [speridlabs/iris-3b](https://huggingface.co/speridlabs/iris-3b) · 2026-10-05 · apache-2.0 Iris-3B Turbo, in the turbo/ folder of this repo, is a 4-step distilled version of Iris-3B. 3B params. License: apache-2.0. Runs on: need Q4 3.2 GB / Q8 4.7 GB / FP16 7.5 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060) - [perplexity-ai/pplx-decider-v1.1-27b](https://huggingface.co/perplexity-ai/pplx-decider-v1.1-27b) · 2026-10-05 · apache-2.0 pplx-decider-v1.1-27b is an update of our pplx-decider-v1-27b model, that boost the Decision Index score from 56.4 to 61.56, now outperforming Jev by more than 3.5 points, while using the same backbone. Dense, 26B params. Context: 256K tokens (config). License: apache-2.0. Vendor-reported: "pplx-decider-v1.1-27b is an update of our pplx-decider-v1-27b model, that boost the Decision Index score from 56.4 to 61.56, now outperforming Jev by more than 3.5 points, while using the same backbone." Runs on: need Q4 18.4 GB / Q8 30.3 GB / FP16 55.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [LiquidAI/d1-omni-600M](https://huggingface.co/LiquidAI/d1-omni-600M) · 2026-10-05 · lfm1.0 d1-omni-600M is a 600M parameter decision model built on LFM2.5-Encoder-350M. Dense, 587M params. Context: 125K tokens (config). License: lfm1.0. Runs on: need Q4 1.5 GB / Q8 1.7 GB / FP16 2.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [inclusionAI/LemonDecider](https://huggingface.co/inclusionAI/LemonDecider) · 2026-10-05 · apache-2.0 LemonDecider 4B is a typed decision engine: it reads a state and a question with a list of options, runs one forward pass, and returns a probability distribution over the options (choice / noul questions) or an ordered rating (score questions). Dense, 4.5B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 4.8 GB / Q8 6.9 GB / FP16 11.2 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [amazon/ALoDLM-8B](https://huggingface.co/amazon/ALoDLM-8B) · 2026-10-02 · cc-by-nc-4.0 ALoDLM is a family of diffusion language models developed at Amazon, available as ALoDLM-1.7B and ALoDLM-8B and initialized from the corresponding Qwen3 backbones. Dense, 8.2B params. Context: 40K tokens (config). License: cc-by-nc-4.0. 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) - [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) - [nvidia/Flora-Rewarder](https://huggingface.co/nvidia/Flora-Rewarder) · 2026-10-02 · nvidia-license Training labels are a combination of human-written annotations, automatic success signals recorded with the data, and VLM-generated annotations. Dense, 28B params. Context: 256K tokens (config). License: nvidia-license. 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) - [amazon/ALoDLM-1.7B](https://huggingface.co/amazon/ALoDLM-1.7B) · 2026-10-02 · cc-by-nc-4.0 ALoDLM is a family of diffusion language models developed at Amazon, available as ALoDLM-1.7B and ALoDLM-8B and initialized from the corresponding Qwen3 backbones. Dense, 2B params. Context: 40K tokens (config). License: cc-by-nc-4.0. Runs on: need Q4 3.1 GB / Q8 4.1 GB / FP16 6.0 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [fishaudio/drama-3](https://huggingface.co/fishaudio/drama-3) · 2026-10-01 · fish-audio-research-license Fish Audio Drama 3 is Fish Audio's new-generation text-to-speech model, built for multi-character scenes and long-form content. Mixture-of-experts, 36B total, 3B active (256 experts, 8 per token). Context: 256K tokens (config). License: fish-audio-research-license. Runs on: need Q4 25.5 GB / Q8 40.2 GB / FP16 73.3 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [canberkkkkkk/ema-lightning](https://huggingface.co/canberkkkkkk/ema-lightning) · 2026-10-01 · apache-2.0 Text to speech model from canberkkkkkk (HF pipeline tag; the card has no one-line description). License: apache-2.0. - [perplexity-ai/pplx-decider-v1-27b](https://huggingface.co/perplexity-ai/pplx-decider-v1-27b) · 2026-10-01 · apache-2.0 pplx-decider-v1-27b is a decision model fine-tuned from Qwen3.8-27B. Dense, 26B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 18.8 GB / Q8 30.8 GB / FP16 55.3 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 32GB GPU (5090), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [nvidia/PixelUMM](https://huggingface.co/nvidia/PixelUMM) · 2026-10-01 · nvidia-one-way-noncommercial-license PixelUMM is an NVIDIA-developed, encoder-free unified multimodal model for joint understanding and generation across text, images, and video directly in pixel space. License: nvidia-one-way-noncommercial-license. - [nvidia/Flora-Captioner](https://huggingface.co/nvidia/Flora-Captioner) · 2026-10-01 · nvidia-license Image text to text model from nvidia (HF pipeline tag; the card has no one-line description). Dense, 27B params. Context: 256K tokens (config). License: nvidia-license. 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) - [ibm-granite/ensemble-forecasting-with-ibm-granite-time-series](https://huggingface.co/ibm-granite/ensemble-forecasting-with-ibm-granite-time-series) · 2026-10-01 - [Cloudflare/clef](https://huggingface.co/Cloudflare/clef) · 2026-09-30 · apache-2.0 Clef is a 27B multimodal model that turns a state and a schema of typed questions into decisions. Dense, 27B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 20.8 GB / Q8 31.9 GB / FP16 57.2 GB; smallest fit Q4: 24GB GPU (3090 / 4090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [Cloudflare/clef-flash](https://huggingface.co/Cloudflare/clef-flash) · 2026-09-30 · apache-2.0 Clef-Flash is a 9B multimodal model that turns a state and a schema of typed questions into decisions. Dense, 9.4B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 8.1 GB / Q8 11.8 GB / FP16 20.3 GB; smallest fit Q4: 12GB GPU (RTX 3060), Q8: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080), FP16: 24GB GPU (3090 / 4090) - [Cactus-Compute/whistle](https://huggingface.co/Cactus-Compute/whistle) · 2026-09-30 · apache-2.0 Whistle is a log-mel front end and a convolutional stem feeding an audio encoder, read by a Needle-shaped decoder through gated cross attention at every layer. License: apache-2.0. - [KittenML/kitten-tts-2](https://huggingface.co/KittenML/kitten-tts-2) · 2026-09-30 · stellon-labs-community-license Text to speech model from KittenML (HF pipeline tag; the card has no one-line description). License: stellon-labs-community-license. - [speach1sdef178/MiniMax-H3-X2-Detail-VAE](https://huggingface.co/speach1sdef178/MiniMax-H3-X2-Detail-VAE) · 2026-09-30 · other Image to video model from speach1sdef178 (HF pipeline tag; the card has no one-line description). License: other. - [facebook/meta-encoder](https://huggingface.co/facebook/meta-encoder) · 2026-09-30 · apache-2.0 Feature extraction model from facebook (HF pipeline tag; the card has no one-line description). 30B params. License: apache-2.0. Runs on: need Q4 23.2 GB / Q8 35.1 GB / FP16 63.1 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [BAAI/AREX-2](https://huggingface.co/BAAI/AREX-2) · 2026-09-29 · apache-2.0 AREX-2 is a 27B-parameter long-horizon agent model from the Beijing Academy of Artificial Intelligence (BAAI). Dense, 27B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 20.3 GB / Q8 31.8 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) - [microsoft/AesCode-8B](https://huggingface.co/microsoft/AesCode-8B) · 2026-09-29 · apache-2.0 Image text to text model from microsoft (HF pipeline tag; the card has no one-line description). Dense, 8.8B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 7.4 GB / Q8 10.9 GB / FP16 18.6 GB; smallest fit Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) - [microsoft/AesCode-32B](https://huggingface.co/microsoft/AesCode-32B) · 2026-09-29 · apache-2.0 Image text to text model from microsoft (HF pipeline tag; the card has no one-line description). Dense, 33B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 23.6 GB / Q8 37.9 GB / FP16 68.6 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [internlm/AdvancedMathBench-AutoVerifier](https://huggingface.co/internlm/AdvancedMathBench-AutoVerifier) · 2026-09-29 AutoVerifier is a learned grader, not a formal proof checker, and can make errors. Mixture-of-experts, 36B total (256 experts, 8 per token). Context: 256K tokens (config). Runs on: need Q4 23.3 GB / Q8 39.9 GB / FP16 74.7 GB; smallest fit Q4: 32GB GPU (5090), Q8: 48GB GPU (RTX 6000 Ada / A6000), FP16: 96GB GPU (RTX Pro 6000 Blackwell) - [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) - [Aratako/Irodori-TTS-v4-Large](https://huggingface.co/Aratako/Irodori-TTS-v4-Large) · 2026-09-27 · gemma Irodori-TTS-v4-Large is a Japanese Text-to-Speech model based on a Rectified Flow Diffusion Transformer (RF-DiT) architecture. 3.3B params. License: gemma. Vendor-reported: "v4-Large achieves higher mean CAM++ cosine similarity at all four reference lengths, although 120-second top-1 accuracy is slightly lower." Runs on: need Q4 3.5 GB / Q8 5.4 GB / FP16 8.5 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 12GB GPU (RTX 3060) - [XiaomiMiMo/MiMo-V2.6-Flash-MOPD](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD) · 2026-09-27 · mit Text generation model from XiaomiMiMo (HF pipeline tag; the card has no one-line description). 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 - [XiaomiMiMo/MiMo-V2.6-Pro-MOPD](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD) · 2026-09-27 · mit Text generation model from XiaomiMiMo (HF pipeline tag; the card has no one-line description). 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 2052.9 GB - [internlm/Intern-Decision-4B](https://huggingface.co/internlm/Intern-Decision-4B) · 2026-09-26 · apache-2.0 Intern-Decision-4B is a multimodal structured decision model fine-tuned from Qwen3.5-4B. Dense, 4.5B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 4.8 GB / Q8 6.6 GB / FP16 11.2 GB; smallest fit Q8: 8GB GPU (4060 / 3070), FP16: 16GB GPU (5060 Ti / 4060 Ti / 4080 / 5080) - [internlm/Intern-Decision-0.8B](https://huggingface.co/internlm/Intern-Decision-0.8B) · 2026-09-26 · apache-2.0 Intern-Decision-0.8B is a multimodal structured decision model fine-tuned from Qwen3.5-0.8B. Dense, 853M params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 1.9 GB / Q8 2.2 GB / FP16 3.1 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [internlm/Intern-Decision-2B](https://huggingface.co/internlm/Intern-Decision-2B) · 2026-09-26 · apache-2.0 Intern-Decision-2B is a multimodal structured decision model fine-tuned from Qwen3.5-2B. Dense, 2.2B params. Context: 256K tokens (config). License: apache-2.0. Runs on: need Q4 2.7 GB / Q8 3.4 GB / FP16 5.8 GB; smallest fit FP16: 8GB GPU (4060 / 3070) - [Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview](https://huggingface.co/Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview) · 2026-09-25 · minimax-h3-community Veda is a learned sparse-attention method for video diffusion models. Dense. License: minimax-h3-community. Vendor-reported: "Visual and audio quality on par with full attention while skipping 90% of it." - [perplexity-ai/pplx-embed-v2-context-9b-preview](https://huggingface.co/perplexity-ai/pplx-embed-v2-context-9b-preview) · 2026-09-25 · mit Feature extraction model from perplexity-ai (HF pipeline tag; the card has no one-line description). Dense, 8.4B params. Context: 256K tokens (config). License: mit. Runs on: need Q4 7.1 GB / Q8 11.0 GB / FP16 18.9 GB; smallest fit Q8: 12GB GPU (RTX 3060), FP16: 24GB GPU (3090 / 4090) ## Runtimes and harnesses (17) - [llama.cpp b11869](https://github.com/ggml-org/llama.cpp/releases/tag/b11869) · 2026-10-24 jinja : fix select and selectattr argument handling (#31059) - [Ollama v0.40.11](https://github.com/ollama/ollama/releases/tag/v0.40.11) · 2026-10-24 Ollama can now generate speech on macOS. - [TensorFold v1.1.4](https://github.com/ashhart/TensorFold/releases/tag/v1.1.4) · 2026-10-24 1.1.4 makes Qwen3.8-27B's cold prompts on CUDA up to five times faster, brings Gemma 4 12B and 31B to the native CUDA engine with their drafters, lets several Flash Next streams verify their drafts in one pass on CUDA, and adds idle unload, - [Hermes Agent v0.22.5](https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.5) · 2026-10-23 Also in the window, undocumented here on purpose: Nous sign-in reworked (every sign-in is the plain device login, the free-tier guest hands over to your account, the result shows who signed in, connectors offer sign-in instead of opening a - [Unsloth v0.1.910-beta](https://github.com/unslothai/unsloth/releases/tag/v0.1.910-beta) · 2026-10-23 Recursive's Flint, MOSS-Image and NVIDIA's Nemotron Bolt 4B are here! - [OpenCode v1.19.4](https://github.com/anomalyco/opencode/releases/tag/v1.19.4) · 2026-10-22 **Advisory:** Security incident Oct 17, 2026: rotate provider keys saved in OpenCode Console ([details](https://ctxsec.com/items/opencode-console-provider-key-breach-2026-10)) - [TensorRT-LLM v1.4.0rc1](https://github.com/NVIDIA/TensorRT-LLM/releases/tag/v1.4.0rc1) · 2026-10-22 - [ExLlamaV3 v1.6.4](https://github.com/turboderp-org/exllamav3/releases/tag/v1.6.4) · 2026-10-20 - [SGLang v0.5.22](https://github.com/sgl-project/sglang/releases/tag/v0.5.22) · 2026-10-19 SGLang now requires Python 3.11 or newer. - [ComfyUI v0.42.1](https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.42.1) · 2026-10-19 Fix some crashing issues a few users experienced when running the MOSS image template workflow. - [vLLM v0.32.0](https://github.com/vllm-project/vllm/releases/tag/v0.32.0) · 2026-10-18 This release features 814 commits from 363 contributors (121 new)! - [MLX-LM v0.32.2](https://github.com/ml-explore/mlx-lm/releases/tag/v0.32.2) · 2026-10-16 - [Open WebUI v0.12.3](https://github.com/open-webui/open-webui/releases/tag/v0.12.3) · 2026-10-16 - [llm-scaler (Intel Arc) vllm-1.5](https://github.com/intel/llm-scaler/releases/tag/vllm-1.5) · 2026-10-16 - [MLX v0.32.4](https://github.com/ml-explore/mlx/releases/tag/v0.32.4) · 2026-10-14 - [KoboldCpp v1.123.1](https://github.com/LostRuins/koboldcpp/releases/tag/v1.123.1) · 2026-10-13 Download and run the koboldcpp.exe (Windows) or koboldcpp-linux-x64 (Linux), which is a one-file pyinstaller for NVIDIA GPU users. - [OpenVINO 2026.4.1](https://github.com/openvinotoolkit/openvino/releases/tag/2026.4.1) · 2026-10-01 You can find the OpenVINO™ toolkit 2026.4.1 release here: