Abliterated models, served over an API.

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.

What you can call right now

None yet, and saying otherwise would waste your time. The abliterated builds live on serverless hosts that carry the long tail of community weights rather than on the mainstream routers, so reaching them is a matter of which upstream is connected, not of whether the models exist. There are a few hundred published.

What is servable today is the uncensored set: 14 loosely aligned finetunes that refuse far less than a frontier model, though they can still decline.

Calling one

Anything that speaks the OpenAI chat completions format works unchanged. Point it at MagmaRouter, use your key, and name the model.

curl https://api.magmarouter.com/v1/chat/completions \
  -H "Authorization: Bearer rl-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthracite-org/magnum-v4-72b",
    "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="rl-...",
)

resp = client.chat.completions.create(
    model="anthracite-org/magnum-v4-72b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

FAQ

Read further