Reducing refusals is not a free capability upgrade: it changes the model’s safety behavior while leaving its underlying abilities largely intact. This release makes that tradeoff available in a locally runnable format, so it is best understood as a research artifact—not a safer or more capable general-purpose assistant.
Key Capabilities
- The underlying 763B-parameter mixture-of-experts model has 384 routed experts, with about six active per token, plus a 197B-parameter Engram lookup table. The Engram data is read from disk on demand, easing RAM pressure but making storage and memory planning central to running it.
- The recommended Mix-256G build uses 235.8 GB in memory and records 93.74% top-token agreement with the unquantized model on the reported held-out evaluation. Mix-128G uses 114 GB, but its higher perplexity and 78.50% agreement make it a substantially degraded fallback rather than an equivalent compact version.
- The abliterated checkpoint’s weighted harmful-prompt refusal score is reported as 0.074, compared with 0.874 for the base model; its benign-prompt over-refusal score is 0.006 versus 0.069. These results describe a sharp safety-behavior change, not a broad capability evaluation: MMLU is the only capability benchmark reported, and the measurements use the unquantized checkpoint.
- The GGUF files target llama.cpp support for the model architecture and are text-only. The advertised million-token context is inherited from the base model; the reported evaluations were conducted at 2K context, so they do not establish long-context performance.
Who It’s For—and the Tradeoff
A fit for researchers studying refusal mechanisms, robustness, and controlled red-team evaluation who can provide substantial memory and add their own safeguards. Look elsewhere for an ordinary assistant, production deployment without independent moderation, or dependable long-context results: the reduced guardrails increase misuse risk, and the smaller build loses considerably more fidelity.
