POST /v1/chat/completions
Qwen 3.8 27B is available as an API preview. The current GPU trial runs until October 16, 2026, 12:20 UTC. All projects pay $0.65 / 1M input and $13 / 1M output tokens. Read availability, model IDs, limits and rates from GET /v1/models.
This synchronous API is separate from image/video jobs. Supply a Bearer API key with generations:write and a unique Idempotency-Key. It supports a subset of Chat Completions, not Responses.
Supported: messages, max_tokens (default 2048, ceiling 4096), temperature, stream, function tools, tool_choice auto/none/required and response_format json_schema. Pilot context: 81920 tokens. Up to 8 image_url and 1 video_url parts in user messages. JPEG/PNG/WebP: up to 32 MiB and 16 MP; MP4/MOV: up to 32 MiB, 1080p, 30 seconds; 48 MiB total. Only approved HTTPS storage hosts (initially creator-app-s3-bucket.s3.eu-central-1.amazonaws.com; contact support for your storage) are accepted; media_asset, arbitrary URLs, audio, n and provider options are unsupported.
Initial limits: 4 running + 8 waiting globally, 2 running + 2 waiting per project; queue wait up to 20 seconds, inference up to 300 seconds. A full queue returns HTTP 429 and Retry-After without reserving funds. Queue expiry releases the hold.
A context-bound maximum is reserved before queue admission. Actual input/output tokens are charged after completion and the remainder released. Input includes text, templates, tools and visual tokens; output includes reasoning. Rates are snapshotted at reservation. Read usage and gensta_billing for final accounting.
Same-key retry returns 409 while queued/running, replays a completed result without a new charge, or returns 502 after failure. Changed payload: 409. Responses are retained for 24 hours, then 410; a key never starts a second paid inference. Use a new key after a terminal failure. Disconnecting after inference starts does not cancel a successful completion or its charge.
With stream=true, SSE final usage and [DONE] follow durable settlement. A mid-stream error is an error event without [DONE]; repeat the same key to resolve the outcome. Before streaming: 400 invalid input, 401/403 access, 402 balance, 429 capacity, 502 model failure, 503 serving disabled.
{"model":"qwen-3.8-27b","messages":[{"role":"user","content":"Suggest three ideas for a product video"}],"max_tokens":256,"stream":true}