./models --sort=newest
Local Models
What I've built and run on my own GPUs, mostly Tesla V100 32GB cards that the big engines stopped caring about. Each page has the downloads, the full recipe, how to get it running, and what we measured along the way.
Weights live on Hugging Face, engine code and launch profiles on GitHub.
V100 Builds
Stock models packed to fit 32 GB Volta cards, plus the inference stack to run them. Sized for 2, 4, or 8 GPUs.
DeepSeek-V4.1-Flash-1017 EXL3
DeepSeek-V4.1-Flash at its native 1M context on 8× V100. EXL3 experts, KL 0.0358 vs the fp32 reference.
GLM-5.3-Flash EXL3
A 321B MoE at ~2.1 bpw on four V100s. 91 tok/s single stream with MTP, or 1M context on eight.
Qwen3.8-27B V100 Quants
The whole 27B VLM packed to 19.5 GB, vision and MTP head intact, on a pair of 32 GB V100s.
Uncensored Models
Abliterated Qwen builds. Refusal directions removed, KL kept low, then quantized for V100, Blackwell, llama.cpp, and Apple Silicon.
Qwen3.8-27B Uncensored
Two de-refused Qwen3.8-27B builds, broad and cyber-targeted, quantized for V100, Blackwell, llama.cpp and Apple Silicon.
Qwen3.6-27B Uncensored
Heretic-abliterated Qwen3.6-27B, two refusal profiles (Cyber, Aggressive), quantized W4A16 and NVFP4 for V100.
Qwen3.6-40B Deckard Uncensored
W4A16 AWQ quant of DavidAU's 40B Opus-distilled Deckard checkpoint, sized for 2× V100, plus a Hermes-calibrated tool-use build.