NeverNO

Qwen 3.8 27B

Live in the free chat

An open model with its refusal reflex taken out. It answers like a capable assistant that doesn't lecture you.

Model facts

Base model
Qwen 3.8, 27B parameters, open weights
Freed with
Aggressive, directional ablation
Runs on
One sealed machine, NVIDIA GPU in confidential mode
Refusal rate
Measured on launch day and published here
Free trial
5 messages, 600 characters in, about 500 tokens out
Memory
None. Each conversation ends when you close the tab

How its refusals were removed

  1. Start from an open model

    Qwen 3.8 27B from Qwen: open weights that anyone can download and inspect.

  2. Find the refusal direction

    Language models learn a single internal direction that pushes them toward "I can't help with that". Heretic measures it on prompts the model refuses.

  3. Remove it from the weights

    That direction is taken out of the model itself. No jailbreak prompt, no filter to slip past: the reflex is gone.

  4. Keep the rest intact

    Heretic keeps the changed model as close as possible to the original on ordinary prompts, so it stays as capable as before.

The one line

Nothing sexual involving minors, ever. This isn't left to the model: our server checks every question and every answer, and blocks it before it reaches you. Everything else is between you and the model.

Which model is next

Open models improve every month. Each milestone unlocks the next one to be freed and sealed, and the attestation will carry the fingerprint of its exact weights.

Coming

Ask it what others won't answer.

Try it free