SubMaroon/Kitchoon-V3-26B-A4B
just landed · new, unvettedmodel
New, unvetted: created or first seen in the last few days, not from a trusted org and not past the signal threshold yet.
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.
- Org
- SubMaroon
- Released
- 2026-10-20
- Status
- just landed · new, unvetted
- License
- apache-2.0 (permissive)
- Parameters
- 25.8B total, 4.0B active
- Architecture
- MoE · 128 experts, 8/token · Gemma4ForConditionalGeneration
- Context
- 262,144 tokens
- Signal
- 9 HF likes · 113 downloads
Runs on
real quant files; need = weights + FP16 KV cache for 8K tokens + 1GB. A sizing calculation, not tested.
- Q4 ~19.8 GB: 24GB GPU (3090 / 4090), 24GB GPU (7900 XTX), Mac 36GB, CPU only, 32GB RAM (slow)
- Q8 ~29.9 GB: 32GB GPU (5090), 48GB GPU (RTX 6000 Ada / A6000), Arc B70 32GB, Mac 48GB, Mac 64GB (M1 Max)
- FP16 ~54.6 GB: 96GB GPU (RTX Pro 6000 Blackwell), 1× DGX Spark (128GB), 2× DGX Spark (256GB), Strix Halo 128GB unified (~96GB usable), Mac 96GB, Mac 128GB, Mac 192GB, Mac 256GB, Mac 512GB, CPU only, 64GB RAM (slow), CPU only, 128GB RAM (slow), CPU only, 256GB RAM (slow)
Links
First published here 2026-10-20 · updated 2026-10-22 · markdown · JSON