Abliterated modelson one API

Abliterated builds of Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017 and MiMo V2.6 Flash, served on one OpenAI-compatible endpoint with tool calling and a million tokens of context. No GPU to rent and no weights to download: you send a request and pay for the tokens.

What abliteration is

A language model that refuses is not consulting a rulebook. Researchers found in 2024 that refusal is carried largely by a single direction in the model's internal activations: push the activations along it and the model refuses harmless requests, remove it and the model stops refusing. Abliteration is the second of those. The direction is measured by comparing how the model responds to requests it refuses and requests it accepts, and it is then projected out of the weights.

The result is a model that behaves the same as before on ordinary work and no longer declines. Because the change is in the weights, it does not rely on a system prompt and it cannot be talked back into refusing by something in a file or a tool result.

Abliterated and uncensored are not the same word

"Uncensored" describes what a model does: it answers. "Abliterated" describes how one kind of uncensored model was made. We keep the two apart in the listing. A model is marked abliterated here only when that is how the build was produced, and the API says so too: every entry in /v1/models carries an abliterated field that is true or false.

shell

curl https://uncensoredopencode.com/v1/models

The abliterated models

Kimi K3.1 ProThe strongest model on the line DeepSeek V4.1 Flash 1017Abliterated DeepSeek for agent loops MiMo V2.6 FlashAbliterated MiMo with vision

The whole line

Model	Build	Context	Tools	Images	Per million tokens
GLM 5.3 Flash `glm-5.3-flash-uncensored`	Uncensored	1M tokens	Yes	Yes	$0.30 in / $1.00 out
Kimi K3.1 Pro `kimi-k3.1-pro-uncensored`	Abliterated	1M tokens	Yes	Yes	$7.50 in / $37.50 out
GLM 5.3 `glm-5.3-uncensored`	Uncensored	1M tokens	Yes	No	$2.00 in / $5.00 out
DeepSeek V4.1 Flash 1017 `deepseek-v4.1-flash-1017-uncensored`	Abliterated	1M tokens	Yes	Yes	$1.25 in / $2.50 out
MiMo V2.6 Flash `mimo-v2.6-flash-uncensored`	Abliterated	1M tokens	Yes	Yes	$1.00 in / $2.00 out
Qwen3.8 27B `qwen3.8-27b-uncensored`	Uncensored	1M tokens	Yes	Yes	$1.50 in / $5.00 out
GLM 5.3 Flash Cyber `glm-5.3-flash-cyber`	Security-tuned	1M tokens	Yes	Yes	On the models page

Call one

shell

curl https://uncensoredopencode.com/v1/chat/completions \
  -H "Authorization: Bearer $UNCENSORED_OPENCODE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "kimi-k3.1-pro-uncensored", "messages": [{"role": "user", "content": "Hello"}]}'

Use them from an agent

opencodeOne provider block, every model, real context limits.Hermes AgentA custom endpoint in config.yaml.Claude CodeThree environment variables.CodexA provider in config.toml.SillyTavernChat Completion with a custom source.

Questions

Why use an API instead of running abliterated weights myself?

Running a large model with a million tokens of context needs far more memory than a desktop GPU has. The abliterated weights people run at home are small or heavily quantised. An API gives you the full-size model and you pay only while you use it.

Does abliteration make a model less capable?

It can shift quality a little in either direction, and it depends on the build. We publish the base model's benchmark score on each model page and say plainly that it is the base model's figure.

Do abliterated models still call tools?

Yes. Every model on this line supports tool calling in the OpenAI format.

Bring your agent.

Get an API key Set up opencode See the models