Models

Choose a model via the model field in your request. Each model runs on dedicated GPU hardware in Sweden.

All chat models support streaming and tool calling, and all of them read images. The thinking aliases (qwen3.6:35b-a3b-thinking, qwen3.5:9b-thinking) are described under Chat completions.

The models are open weights that we host on our own hardware. KB-Whisper is the only model we run that was trained in Sweden.

List models programmatically

curl https://api.staik.se/v1/models \
  -H "Authorization: Bearer sk-st-your-key"

{
  "data": [
    {
      "id": "qwen3.6:35b-a3b",
      "object": "model",
      "owned_by": "staik",
      "context_window": 262144,
      "capabilities": ["chat", "vision", "tools", "systemone"]
    }
  ]
}

The response includes context_window and capabilities per model — useful for clients that configure themselves. systemone means the model can be used with /v1/systemone.

Version-neutral aliases

Want to follow the latest version without hardcoding a version number? Use an alias as model:

When we upgrade a model, the alias follows automatically — your code stays the same. The aliases work on both /v1/chat/completions and /v1/messages.

Thinking only works on /v1/chat/completions. Add -thinking (qwen-35b-latest-thinking, qwen-9b-latest-thinking) to enable thinking. /v1/messages does not support thinking: there the -thinking names are treated as an unknown model and are answered by the default model (qwen3.6:35b-a3b) without thinking.

The model field in the response differs between the endpoints:

• /v1/chat/completions shows the actual model that answered, e.g. qwen3.6:35b-a3b.
• /v1/messages usually shows staik's internal model name: qwen3.5:35b-a3b for qwen-35b-latest (the same model as qwen3.6:35b-a3b) and qwen3.5:9b for qwen-9b-latest. Streamed requests with tools show the actual model instead. If you need the exact version, read actual_model in GET /v1/models.

GET /v1/models lists the aliases with alias: true and actual_model, which shows the version the alias currently points to (the -thinking variants are not listed):

{
  "id": "qwen-9b-latest",
  "object": "model",
  "owned_by": "staik",
  "alias": true,
  "actual_model": "qwen3.5:9b",
  "context_window": 262144,
  "capabilities": ["chat", "tools", "systemone"]
}

Want predictable behavior over time? Use the full model name instead, e.g. qwen3.6:35b-a3b.

Your own GPUs (BYO): unmetered lane

If your organization has connected its own GPUs (Your GPUs), you choose the lane on the model name:

• Prefixed — <org_slug>/<model> runs unmetered against your own hardware (no daily/hourly limit, no token deduction)
• A bare model name goes to staik's shared pool and is metered against your plan

# Unmetered: run against the organization's own GPU (prefix = your org slug)
curl https://api.staik.se/v1/chat/completions \
  -H "Authorization: Bearer $STAIK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "acme/qwen3.8:27b", "messages": [{"role": "user", "content": "Hi"}]}'

If your org has no healthy node serving the model you get a 503 — no silent fallback to staik. GET /v1/models lists your prefixed models. The prefix works the same on /v1/chat/completions and /v1/messages.