Removing a model’s refusal behavior is not a cost-free switch: this edit eliminates refusals on the tested harmful prompts, but also changes some ordinary responses and lowers MMLU performance. That makes the key question not just what it will answer, but how much behavior changes along the way.
Key Capabilities
- The rank-1 LoRA edits 49 residual-stream writers across the model, targeting layer-specific refusal directions rather than retraining the full model. This keeps the modification distributable as a small adapter as well as merged GGUF files.
- The underlying 26.4B-parameter MoE has 3.4B active parameters per token and accepts text and image input. Ready-made quantizations range from Q4_K to IQ2_M, allowing different storage and quality tradeoffs for local inference.
- In the reported tests, the adapter reduced refusals on harmful prompts to zero across the tested English and Russian sets, and reduced safe-prompt refusals on XSTest from 4% to zero. These are specific benchmark results, not a guarantee of behavior on other prompts.
- The measured costs matter: MMLU dropped from 80.65% to 79.15%, and first-token KL on harmless prompts was 0.94 against a 0.10 threshold. The model therefore does not preserve the base model’s behavior exactly, even where refusal is not the goal.
Who It’s For
A good fit for researchers and experienced local-model users evaluating refusal editing, multilingual behavior, or GGUF inference. Look elsewhere if you need the base model’s refusal behavior or close behavioral fidelity: the reported MMLU regression is statistically significant, and the first-token divergence exceeds the project’s own threshold. The model may answer requests the original declines.