MiniMax M2.7 abliterated weights hit 4% refusal rate

MiniMax M2.7 ultra uncensored heretic, a safety-abliterated fine-tune from llmfan46, refused only 4 out of 100 prompts on an internal safety benchmark. The weights dropped this week on HuggingFace in both BF16 safetensors and quantized GGUF formats, targeting practitioners who run models locally without server-side filters.

The 4-out-of-100 refusal rate places this checkpoint among the more permissive abliterated releases in the open-weight scene. Most base models from major labs refuse 80 to 100 percent of adversarial prompts by design. Abliteration—a fine-tuning technique that removes refusal behavior while attempting to preserve general capabilities—has become standard practice in the uncensored model community, with practitioners sharing checkpoints across model families from Llama to Qwen to MiniMax. The abliteration process here produced a KL divergence of 0.0452 from the base MiniMax M2.7 model, a conservative shift that suggests the model retains most of the original reasoning and language structure while stripping safety guardrails.

The GGUF repo includes standard quantization levels for CPU and consumer GPU inference, making the model accessible to users without high-end hardware. GGUF has become the de facto format for local inference tooling like llama.cpp and Ollama, and pre-quantized weights eliminate the need for users to run conversion scripts themselves. llmfan46 has released a series of abliterated checkpoints across multiple model families; the M2.7 heretic weights are licensed under the same terms as the original MiniMax release.