Models

XField serves a short list of uncensored open-weight models from DeepSeek, Moonshot AI, Z.ai, and Qwen. Every model works on every protocol: Chat Completions, Responses, and Messages. All of them stream, call tools, and can reason before they answer. Pass the id exactly as shown below in the model field.

DeepSeek V4.1 Flash, Kimi K3.1 Pro, GLM 5.3 Flash and Qwen 3.8 27B read images in the user message, up to 4 per message; GLM 5.3 is text only. Kimi K3.1 Pro also reads video, sent as a base64 video_url part on Chat Completions only. Every model answers in text. Every request names its model with model; there is no default.

Which model should I use?

The short version:

• Most work, including coding agents and security research? Use DeepSeek V4.1 Flash. It leads our models on reasoning, coding and cybersecurity benchmarks, and it is cheap once your system prompt and tools are cached.
• Hard agentic coding, research, or very long outputs? Use Kimi K3.1 Pro, Moonshot AI's flagship, with up to 922K tokens of output. It also reads video. It costs more than the others, so save it for work that needs it.
• Long outputs, or a second opinion? Use GLM 5.3. Its 128K output limit covers most long reports, and a different model family catches different mistakes.
• Generating or labeling a lot of data? Use GLM 5.3 Flash, the lighter GLM, built for batch work.
• Quick chat or drafts? Use Qwen 3.8 27B. It is small and answers fast with thinking off.

You can switch models per request, so it is common to plan with GLM 5.3 and execute with DeepSeek V4.1 Flash in the same tool.

Listing models

GET /api/v1/models returns every model you can call, in the familiar OpenAI list shape with a few extra fields. Most clients that support model discovery pick this up automatically.

curl https://xfield.cloud.brutal.ai/api/v1/models \
  -H "Authorization: Bearer $XFIELD_API_KEY"

A trimmed response:

{
  "data": [
    {
      "id": "deepseek-v4.1-flash-1017-uncensored",
      "name": "DeepSeek V4.1 Flash 1017",
      "context_length": 1048576,
      "architecture": {
        "input_modalities": ["text"],
        "output_modalities": ["text"]
      },
      "supported_parameters": ["messages", "max_tokens", "temperature", "tools", "tool_choice", "reasoning", "stream"],
      "reasoning": {
        "default_effort": "high",
        "effort_mode": "levels",
        "supports_exclude": true
      }
    }
  ]
}

GET /api/v1/models/<id> returns a single entry. The reasoning block tells you whether a model thinks by default and whether it takes effort levels or a plain on/off switch. See Reasoning for the details.

Context and output limits

Context covers everything in the request plus the response. Reasoning tokens count toward the output limit, so leave headroom in max_tokens when thinking is on. If a request does not fit, you get a 400 with a message that says so.