Qwen3.8-27B-heretic-MTP-BF16

An abliterated build of Qwen/Qwen3.8-27B in BF16, with the MTP heads and full vision tower intact. This is the source checkpoint. Quantize it to whatever format you need.

For the FP8 build, the measured numbers, the flags, and the full method writeup, see OptimizeLLM/Qwen3.8-27B-heretic-MTP-FP8.

• Base: Qwen3.8-27B (27.8B dense, 64 layers, hybrid linear/full attention, native image and video).
• Abliteration: HERETIC 1.4.0 with local patches. Two-stage slot-grouped pipeline. KL 0.065 against base on harmless prompts.
• Refusals: one out of 99 for the base model, on a 100-prompt harmful-behaviors set reviewed manually.
• Non-refusing, not neutral. Directional ablation removes the tendency to decline, not a trained-in viewpoint. It will engage with any topic; it can still reflect the base model's lean on one.
• Tokenizer: official Qwen3.8-27B's, byte-identical to upstream.

The reproduce/ folder holds the HERETIC config, seed parameters, and the local patches.

You are responsible for what you do with it.

Thanks

• Qwen Team, for Qwen3.8-27B and the open release.
• p-e-w, for HERETIC.
• The community that first worked out the two-stage slot-grouped MPOA recipe on the Qwen3.6 family.

License

Inherits the base Qwen3.8-27B license. Apache-2.0.