Brutal AI

Uncensored models.

Open models. Refusal layer removed.

We built this for ourselves. Our engineers use these models every day on client projects and internal development, for work like security audits and red-teaming where stock models stop mid-task. We have spare capacity, so a few teams can use it too.

What abliteration does
Stock models learn one internal direction that pushes sensitive requests over a threshold, and the model refuses.Each dot is a prompt. Illustration, not measured data.

The models.

Six models. One API key.

Every text model is our abliterated version of the open model: the same weights, with only the direction that leads to refusals removed. In the API they carry the suffix -uncensored.

ModelBest forInputContextMax outputTokens per secFirst token
Model pageBest forCoding agents, hard reasoningInputText, imageContext1MMax output64KTokens per sec108First token2.1 s
Model pageBest forLong reports, large code changesInputTextContext1MMax output128KTokens per sec92First token2.6 s
Model pageBest forBatch jobs, labelingInputText, imageContext1MMax output32KTokens per sec71First token2.0 s
Model pageBest forAgentic coding, researchInputText, image, videoContext1MMax output922KTokens per sec29First token2.7 s
Model pageBest forChat, quick draftsInputText, imageContext128KMax output32KTokens per sec42First token1.0 s
MiniMax H3MiniMax · Video

Reference-to-video. 1 to 9 reference images, clips of 4 to 15 seconds, up to 1344 × 768, and up to 3 LoRAs per run.

All five text models support tool calls and reasoning. Full specs are in the model catalog.

Qwen answers first. DeepSeek writes fastest.

Each dot is one request. Rings mark the median. Select a model above or in the chart.

045901351800s1s2s3s4s5sTokens per secondTime to first tokenGLM 5.3GLM 5.3 FlashKimi K3.1 ProQwen 3.8 27BDeepSeek V4.1 Flash

Measured 6 Nov 2026: 5 streamed requests per model on shared capacity, up to 400 tokens, same prompt, reasoning tokens included. Your numbers will vary, so test with your own prompts.

Keep your SDK. Change two lines.

OpenAI and Anthropic formats both work, and so do Claude Code and Codex. Quickstart

curl https://xfield.cloud.brutal.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $XFIELD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-1017-uncensored",
    "stream": true,
    "messages": [{"role": "user", "content": "Hello."}]
  }'

What teams use them for.

Red teaming.

Authorized penetration tests and AI red-team evaluations.

Red teaming guide

Coding agents.

Agents that keep working through the security-adjacent parts of a job.

Claude Code setup

Your prompts stay yours.

On the text API, prompts and replies pass through and are not stored or used for training. Each request leaves a usage record with model, token counts, timing and cost.

How we handle data

Guides.

Your first answer.

Ask for access.

Tell Lukas who you are and what you are building. You get an invitation by email.

Create a key.

Sign in and open Settings, then API keys. Each key is shown once.

Change one URL.

Point your SDK or agent at XField, our API, and pick a model. It is OpenAI- and Anthropic-compatible, hosted in the EU, and does not store your prompts.

How you get access.

Shared API.

Use our spare capacity with an API key and pay for what you use. We maintain the models and help when something breaks.

Your own instance.

Capacity reserved for your team, running your choice of models and your own guardrails, with the same operations and support.

Fine-tuning.

Image and video models tuned to your style or products, and small custom models where a large one is too slow or too expensive.

Ask for a key.

Tell Lukas what you are building and where your current model refuses. You get access and pricing back by email.

Access is reviewed company by company. Uncensored models do not change what is legal: you remain responsible for how you use them and for their output, under our acceptable use policy.

Frequently asked questions

What does abliterated mean?

Stock models learn an internal direction that pushes sensitive requests over a threshold, and the model refuses. Abliteration removes that direction. XField's abliterated models have the same weights as the open models, with only that direction removed. In the API they carry the suffix -uncensored.

Which models are available on XField?

Five text models: DeepSeek V4.1 Flash (deepseek-v4.1-flash-1017-uncensored), GLM 5.3 (glm-5.3-uncensored), GLM 5.3 Flash (glm-5.3-flash-uncensored), Kimi K3.1 Pro (kimi-k3.1-pro-uncensored) and Qwen 3.8 27B (qwen-3.8-27b-uncensored). There is also MiniMax H3 for video from reference images. All text models support tool calls and reasoning.

How do I get an API key?

Email Lukas at lukas@brutal.ai and tell him who you are and what you are building. You then get an invitation by email. You create the key under Settings, API keys, and it is shown only once. Brutal AI reviews access company by company.

Who runs XField?

Brutal AI in Vienna. Brutal AI runs the compute for its own systems and offers spare capacity to developers. Its own engineers use these models every day on client projects and internal development.