Abliterated models.

What abliteration actually does

A language model's tendency to refuse is not spread evenly through the weights. It concentrates in a single direction in the residual stream, and that direction can be found by running the model over a batch of harmful prompts and a batch of harmless ones and taking the mean difference of the activations at the last token.

Abliteration projects that direction out of the embedding matrix, every attention output projection and every MLP output projection. What is left cannot represent refusal, so it does not refuse. No retraining, no fine-tuning data, no prompt engineering: it is surgery on the weights, and it survives whatever you put in the context window. The method step by step is on the abliteration explainer.

Abliterated models you can call on MagmaRouter

Every model below is abliterated and on sale today, at the price shown per million tokens.

The uncensored models next to them

The rest of the line is uncensored rather than abliterated: builds with the refusal training taken off, which refuse far less than the stock model but were not edited along the refusal direction. Some are larger and some are cheaper than the abliterated pick, so for many jobs one of them is the better choice.

Who publishes abliterated models

Most abliterated weights are community releases on Hugging Face rather than hosted products. FailSpy published the first ones and named the technique. Maxime Labonne (mlabonne) wrote the tutorial most people learned it from, and released NeuralDaredevil 8B Abliterated, an abliteration followed by a DPO retune to win back what the edit cost. huihui-ai has since published abliterated versions of most major open models. abliteration.ai is the company that sells its own abliterated models as a hosted API.

MagmaRouter does not offer any of those. It sells the line above, and where a model people search for has no equivalent on it, the page for that model says so and names the nearest one: abliteration.ai's Abliterated Model Large V2, for example, or the uncensored pages for GLM, Kimi, DeepSeek, MiniMax and the rest.

Calling one

Anything that speaks the OpenAI chat completions format works unchanged. Point it at MagmaRouter, use your key, and name the model. The same models answer on the Anthropic Messages format too, so Claude Code drives them directly.

curl https://api.magmarouter.com/v1/chat/completions \
  -H "Authorization: Bearer sk-mr-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-1017-xploded",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

From Python

# pip install openai
from openai import OpenAI

client = OpenAI(
    base_url="https://api.magmarouter.com/v1",
    api_key="sk-mr-...",
)

resp = client.chat.completions.create(
    model="deepseek-v4.1-flash-1017-xploded",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

FAQ

Read further