Continue.dev

Last updated:

On this page

Run Continue in VS Code or JetBrains against staik — one model per role: 35b for chat/edit, 9b for autocomplete, and bge-m3 for embeddings (codebase indexing).

Configuration (config.yaml)

yaml
name: staik
version: 0.0.1
schema: v1
models:
  - name: staik qwen3.6 35b
    provider: openai
    model: qwen3.6:35b-a3b
    apiBase: https://api.staik.se/v1
    apiKey: ${{ secrets.STAIK_API_KEY }}
    defaultCompletionOptions:
      contextLength: 131072   # half of 262k — fast + headroom for output
      maxTokens: 8192
    roles:
      - chat
      - edit
      - apply
  - name: staik qwen3.5 9b
    provider: openai
    model: qwen3.5:9b
    apiBase: https://api.staik.se/v1
    apiKey: ${{ secrets.STAIK_API_KEY }}
    defaultCompletionOptions:
      contextLength: 16384    # deliberately low — fast autocomplete
      maxTokens: 4096
    roles:
      - autocomplete
  - name: staik bge-m3
    provider: openai
    model: bge-m3:latest
    apiBase: https://api.staik.se/v1
    apiKey: ${{ secrets.STAIK_API_KEY }}
    roles:
      - embed

Put your key in Continue's secrets (or replace the placeholder with the sk-st key directly).

Context window

Set contextLength to roughly half the model's window. Continue otherwise tends to fill the whole window with repo context, and a full 262K is both slower and eats more of the token budget. If you need really large repo contexts, raise the number again.

ModelFull windowRecommended contextLength
qwen3.6:35b-a3b262,144131,072
gemma4:31b262,144131,072
qwen3.5:9b262,14416,384 (autocomplete)

config.json (legacy format)

Running an older Continue version? The setup is the same:

json
{
  "models": [
    {
      "title": "staik qwen3.6 35b",
      "provider": "openai",
      "model": "qwen3.6:35b-a3b",
      "apiBase": "https://api.staik.se/v1",
      "apiKey": "sk-st-your-key",
      "contextLength": 131072,
      "completionOptions": { "maxTokens": 8192 }
    }
  ],
  "tabAutocompleteModel": {
    "title": "staik qwen3.5 9b",
    "provider": "openai",
    "model": "qwen3.5:9b",
    "apiBase": "https://api.staik.se/v1",
    "apiKey": "sk-st-your-key"
  },
  "embeddingsProvider": {
    "provider": "openai",
    "model": "bge-m3:latest",
    "apiBase": "https://api.staik.se/v1",
    "apiKey": "sk-st-your-key"
  }
}

Indexing stays in Sweden

Continue indexes your whole codebase with the embeddings model. Put bge-m3:latest on the embed role so indexing also stays on Swedish GPU infrastructure instead of hitting a third-party provider.

Want to run Claude Code or OpenCode instead? See Claude Code and OpenCode.