Kilo Code

From the settings#

In VS Code: Settings (gear icon) → Providers → at the bottom, Custom provider.

The list read from /v1/models also holds models that do not chat, such as bge-m3 and whisper-1. Leave those out.

From kilo.jsonc#

The context window, the output limit and the tool flag of each model are set in kilo.jsonc. This is the form Kilo Code's documentation gives for an OpenAI-compatible provider, with AI Tokens's values:

{
  "$schema": "https://app.kilo.ai/config.json",
  "model": "openai-compatible/qwen3.8-27b",
  "provider": {
    "openai-compatible": {
      "options": {
        "apiKey": "{env:API_KEY}",
        "baseURL": "https://api.aitokens.ch/v1"
      },
      "models": {
        "qwen3.8-27b": {
          "name": "Qwen3.8 27B",
          "tool_call": true,
          "limit": { "context": 262144, "output": 8192 }
        },
        "qwen3-coder-30b": {
          "name": "Qwen3-Coder 30B",
          "tool_call": true,
          "limit": { "context": 32768, "output": 8192 }
        }
      }
    }
  }
}

{env:API_KEY} reads the key from the environment variable API_KEY. Keep the key in the global file, in ~/.config/kilo: a kilo.json committed to a project cannot read environment variables.

Which model#

qwen3.8-27b for most work, for its 262,144 tokens of context; qwen3-coder-30b for quick tasks. Both returned tool calls on 11 October 2026, with and without streaming. More in Which model.

Limits#

• output up to 8192. The API refuses a max_tokens above 8,192 with 422.
• No reasoning to show: the API does not return the reasoning of qwen3.8-27b.