MODEL FIELD NOTES
Inside the models.
Four exact Hugging Face artifacts, followed from upstream architecture to post-training, refusal reduction and practical hosting. Every number links back to a primary source; publisher claims stay labeled as publisher claims.
huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated
The multimodal proof of concept: how huihui-ai applied abliteration to a 36-billion-parameter mixture-of-experts checkpoint—and why “uncensored” still needs caveats.
Read article → 35B MoE · Agentic codingYuYu1015/YuYu1015-Ornith-1.0-35B-abliterated
Can self-scaffolding survive abliteration? The story of DeepReinforce's coding model and YuYu1015's corrected weights-only refusal-reduction pass.
Read article → 9B class · Long contexthuihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated
The small model with the very long name: Qwen roots, Claude-derived post-training traces and a “1M” label that deserves a closer look.
Read article → 397B total · W4A16cebeuq/Ornith-1.0-397B-abliterated-W4A16
Surgery on a model too large to hold at once: streaming abliteration and four-bit conversion that still produces nearly 196 GiB of weights.
Read article →