Removing refusals without retraining or changing prompts shifts the safety–capability trade-off into the model weights themselves. That makes this release a useful case study in how refusal behavior can be altered independently of a model’s interface and broad capabilities—but also means the resulting behavior is persistent rather than a runtime setting.
Key Capabilities
- The release patches the native FP8 weights and keeps the upstream FP4 mixture-of-experts components, tokenizer, and prompt-encoding system unchanged. It is a weight-level modification, not a prompt jailbreak or fine-tuned adapter.
- Vision, video, tool calling, reasoning-effort modes, and the 1-million-token context are retained, so the change applies to a general multimodal model rather than a text-only variant.
- The publisher reports 83.77% on its 1,140-question MMLU test, compared with 87.63% for the base model, and 320/320 compliant responses on its HarmBench evaluation. These are publisher-run results; the compliance score is not evidence of safety or factual reliability.
Who It’s For
Great fit for researchers studying refusal behavior, weight editing, and model governance who have suitable high-end FP8 hardware and can evaluate outputs carefully. Look elsewhere if you need built-in refusal safeguards, a smaller model, or a ready-to-run hosted service: this repository is gated, requires substantial GPU resources, and its altered behavior can include responses to harmful requests.
