A Local 9B Model with DeepSeek Distillation and Heretic v2 Goes Viral on Hugging FaceBrief · digest for November 20
Brief≈ 1 min

A Local 9B Model with DeepSeek Distillation and Heretic v2 Goes Viral on Hugging Face

The name DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic looks like someone fell asleep on the keyboard. Even so, the Apache-2.0 checkpoint has racked up 9,664 likes and 38,837 downloads, while its GGUF builds for Ollama and llama.cpp have passed 888,000 downloads.

Two tricks sit inside. The first is distillation: a small model learns from a large model's finished reasoning chains, like a student sitting next to the top scorer in class. The base Qwen3.5-9B was fine-tuned on roughly 250,000 math and hard-science problems from DeepSeek-V4-Pro. Per the author's measurements, it uses 36% fewer reasoning tokens and scores 94.50% on GSM8K and 90.53% on MMLU-Pro across math, physics, and chemistry (up from 89.60% on the original Qwen3.5-9B).

The second step is refusal ablation via Heretic v2: targeted weight edits without retraining that turn off automatic refusals while barely touching knowledge. Refusals dropped from 100/100 to 3/100 with low drift from the original weights (KL divergence of 0.0422).

The catch is simple: all numbers come from the authors without independent auditing, and accuracy was measured before running Heretic v2. Safety filters were stripped on purpose, so this build should not go into a public service without external moderation.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 4