Text generation

Gemma 4 31B

Generate streamed text and structured function calls with Gemma. Enable the model for your project, then use the API key you manage in the Wave dashboard.

Model ID: gemma4-bf16. Endpoint: POST https://api.dotwave.ai/v1/chat/completions.

Start a generation

Create an API key in API keys, keep it on your server and set DOTWAVE_API_KEY.

curl --no-buffer https://api.dotwave.ai/v1/chat/completions \
  -H "Authorization: Bearer $DOTWAVE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Session-Affinity: conversation-example-001" \
  -d '{"model":"gemma4-bf16","messages":[{"role":"user","content":"Say hello in French."}],"stream":true,"max_tokens":80,"reasoning_effort":"none","stream_options":{"include_usage":true}}'

Reuse X-Session-Affinity for every turn of the same conversation, with a different opaque ID for each new call. See the affinity rules.

Responses use server-sent events. Read text from choices[].delta.content, function calls from choices[].delta.tool_calls, and the final usage event before [DONE].

Python, TypeScript, LiveKit and tool-calling guide →

Parameters and limits

  • Text messages and function tools; images, audio and video are not accepted.
  • stream: true is required. Non-streaming requests are not supported.
  • Context: up to 32,768 tokens, including the prompt and generated response.
  • max_tokens or max_completion_tokens: 1–4,096; default 256.
  • Use reasoning_effort: "none" for the voice-agent configuration.

Access hours and billing follow your project agreement. The dashboard shows your configured access hours and recorded usage.

Access and errors

Check GET /v1/models to see whether Gemma is enabled for your project. A valid key outside its service hours returns 403 service_closed. Creating another key preserves the project's access hours.

Unavailable generation returns a non-200 response before output begins. If a stream is interrupted after it begins, treat it as a failed generation. Wave does not retry the generation or select another provider for you.

See error handling and the API reference.