Qwen3-8B ZeroZero Heretic MLX
4-bit MLX build of Qwen3-8B ZeroZero Heretic for Apple Silicon Macs. Quantized with mlx-lm at 4.5 bits per weight.
The results, full limits and method are on the main model page. The numbers there were measured on the full 16-bit model, not this 4-bit build, so expect small differences.
Results in short
| Measure | Change | Before → after |
|---|---|---|
| Refusals on Heretic's harmful-prompt test | 91% fewer | 100/100 → 9/100 |
| Responses hitting the slop list | 68% fewer | 78/100 → 25/100 |
| Slop words per 1,000 | 78% fewer | 16.5 → 3.6 |
| Fact and logic checks passed | No change | 13/15 → 13/15 |
Slop list hits compare against the decensored model, before the noslop pass. The others compare against base Qwen3-8B.
Use it
pip install mlx-lm
hf auth login
mlx_lm.generate --model LastMachine/Qwen3-8B-ZeroZero-Heretic-MLX --prompt "Write the opening of a short story about a lighthouse keeper." --max-tokens 500
Safety warning
This model will answer many requests that a normal model refuses, including harmful ones. Its output can be dangerous, false or biased.
Fewer refusals doesn't mean more truth. On Tiananmen Square it gives a sanitized official account. Don't rely on it for facts about China or for financial, medical or legal decisions. You are responsible for how you use it.
License and credits
Released under Apache 2.0, the license of the base model. Modified with Heretic 2.0.0, based on Qwen/Qwen3-8B. This is an independent project, not affiliated with or endorsed by Qwen, Alibaba Cloud, Anthropic or the Heretic project.