Qwen3.6 heretic variant keeps MTPs

This is mostly a distribution and alignment tweak, not a new base model, but it matters for people trying to run Qwen3.6 locally with fewer refusals and full MTP behavior intact.

• –The model card says the original 19 MTP tensors are preserved; the GGUF count looks different because `gate_up_proj` is split in that format.
• –The format spread is the real value here: safetensors for Transformers users, GGUF for local runners, and NVFP4/GPTQ for lower-memory deployments.
• –The reported KL divergence of 0.0015 suggests the ablation stayed close to the base model while changing refusal behavior materially.
• –Because this is an uncensored fork, users should validate quality and safety on their own workloads before treating it as a drop-in replacement.
• –The benchmark framing is useful, but it is still an author-provided eval, so independent verification would matter for serious adoption.