Qwen3.8 Omni Flash Realtime

Qwen3.8 Omni Flash Realtime handles live text, audio and video-frame inputs with streamed text and speech responses.

A live conversation route

The realtime service is intended for a continuing session rather than a single uploaded-file response. The documented interface supports streamed audio, video frames, interruptions and tool use. Audio input and output use different sample rates, so configure the recorder and player separately. Use the WebSocket domain for the workspace’s region and the supported SDK version.

Session design

A useful assistant needs to handle turn boundaries, interruptions and tool results without losing the user’s intent. Test a user who starts speaking over an answer, changes a request midway and asks a question about a newly supplied frame. Verify that cancelled audio stops playing and that a late tool result does not answer an obsolete request. Compare the delay to the first meaningful response, not just the model’s total generation time.

Tools and limits

The publisher documents function tools and remote MCP connections. Tool use and the built-in web-search option have compatibility restrictions and should not be combined without following the session documentation. The offline Omni model’s context allowance should not be copied into this profile as an established realtime limit. Similarly, the documented output-truncation setting is not a promise that the model will generate exactly that many tokens. Keep session identifiers, audio format and connection errors in your application logs so a failed interaction can be reproduced.

Sources

Publisher documentation · Publisher documentation