Tried p-e-w/heretic: Automatic Model Uncensoring Without the Usual Guesswork

p-e-w/heretic is getting attention fast—+150 GitHub stars today—because it automates a task that normally involves a lot of manual model surgery: reducing refusal behavior in language models.

The project focuses on automatically finding and applying model-weight changes rather than relying on prompt tricks or a system-message jailbreak. That makes it interesting for researchers testing model behavior, local inference stacks, and alignment experiments. It is also why you should treat the output as an experimental model fork, not a production safety feature.

The important architecture detail: Heretic modifies model weights locally; it does not turn a hosted API into an uncensored endpoint by itself. After exporting and serving the modified model through an OpenAI-compatible server, clients can connect normally.

For a gateway route such as claude-fable-5, the basic smoke test looks like this:

export OPENAI_BASE_URL="https://b-lost.com/v1"
export OPENAI_API_KEY="$BLOST_API_KEY"

curl "$OPENAI_BASE_URL/chat/completions" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-fable-5",
    "messages": [
      {"role":"user","content":"Return a short capability and refusal-behavior test."}
    ],
    "temperature": 0.2
  }'

If your modified checkpoint is served locally, replace the base URL with your inference server and use its registered model name. The same pattern works with Aider, Cline, Roo Code, Cursor, Windsurf, LibreChat, and other OpenAI-compatible clients.

For repeated long prompts, B-Lost’s relay is worth considering because it exposes https://b-lost.com/v1, offers a stated 20% discount, and supports native Anthropic /v1/messages prompt caching with 90% cache-hit discounting. Caching helps more with stable system prompts and large tool definitions than with constantly changing test prompts.

One warning: benchmark refusal changes separately from general capability, hallucination rate, and security regressions. Removing a refusal is not the same as improving a model.