Qwen’s September 17–21 releases span translation, audio conversation and multimodal work. Treating them as one interchangeable upgrade would hide the most useful distinction: what the user is trying to do, and how the application exchanges information with the model. Official release log
A recorded meeting summary, a live interpreter and an assistant watching a camera feed have different requirements. Start there before choosing an endpoint.
Four releases to separate
| Release date | Endpoint | Documented role |
|---|---|---|
| September 17 | qwen3.8-livetranslate-flash-realtime | Realtime audio/video translation; 60 source languages and spoken output in 29 languages |
| September 18 | qwen3.8-omni-flash | Text, image, audio and video input; up to 1M context |
| September 20 | qwen-audio-3.1-realtime-plus | Duplex speech conversations, with function calling and web search |
| September 21 | qwen3.8-omni-flash-realtime | Realtime audio/video interaction, text/audio output and remote MCP tools |
These capabilities and dates come from Qwen’s changelog. The realtime Omni entry lists WebSocket, WebRTC and AOQ access. The translation entry describes a WebSocket interface. An API compatibility claim for one endpoint should not be carried across the whole group without checking its documentation. Release details
Evaluate the interaction, not just the answer
For a translation workflow, a useful trial includes names, interruptions, mixed languages and corrections. A fluent sentence can still be wrong in the one detail the listener needs. Compare the output with a reference prepared by someone who understands both languages, and measure delay over complete exchanges.
For a voice assistant, test what happens when the user interrupts a tool call or changes their request halfway through. Record whether the system cancels the action, finishes it, or asks for clarification. The ability to call a function does not decide which of those behaviors your product should permit.
For video analysis, choose a task with a visible ground truth: locating a specific event in a clip, for example. Check temporal accuracy and whether the answer points to evidence a reviewer can inspect. A large context allowance does not establish reliable recall of every moment.
Hosted capabilities and local weights are separate decisions
The release log documents API offerings. It does not establish that these endpoints have matching downloadable checkpoints. Keep that distinction when comparing them with the models in our open-model guide.
Before a pilot, confirm endpoint-specific pricing, regional availability, session limits and data handling. This article does not provide a verified price comparison. Our recommendation is to choose one bounded interaction, define its failure cases, and compare the whole workflow with your current implementation. Better modality support is valuable when it produces a more dependable result for the person using it.