Alibaba Cloud Model Studio's official release table records qwen-audio-3.1-realtime-flash on October 27 for Beijing and the international service through Singapore. It is a duplex conversational voice model with text and audio inputs and outputs. This is a specific cloud service addition, rather than a release of local weights or a demonstrated outcome for every language.
Conversation and memory have different limits
The model page lists a 262,144-token context, while retained audio history has separate limits: up to 50 turns and 300 seconds of cumulative audio. It lists WebSocket, WebRTC and AOQ. These are documented technical parameters, rather than a duration of dependable conversation measured by SIExpo.
The distinction directly affects application design. A large token figure does not mean unlimited access to everything said during a call. An application needs to understand what it keeps in its own history, what it supplies to the model and what has already fallen out of context. Our interpretation is that memory should be evaluated as behavior of the entire service, instead of inferred from a single table entry. Otherwise, an impressive capacity specification can create the wrong expectation about continuity.
An illustrative test would involve a user stating a preference early in a long conversation and changing it later. We would track whether the system uses the latest valid clarification. This is a proposed scenario, not an observed defect or a Qwen test we performed. The same criterion would be useful for any application operating with bounded conversational history.
Deciding when to respond
The guide describes acoustic end-of-turn detection, semantic smart_turn and manual control. In smart_turn, a prerecorded audio sample can support enhancement of a selected speaker. The documentation also describes the requested speech-output language as best-effort rather than guaranteed. We do not present the service as verified for Bulgarian.
Our assessment is that interruptions deserve a separate evaluation. A useful trial would examine short pauses, self-corrections and a second person speaking in the background. Smooth-sounding speech alone cannot establish that the system understood when a speaker's intent was complete. SIExpo has not measured latency, intelligibility or recognition accuracy, and we derive no ranking against other voice models. The meaningful question is whether the actual conversation remains clear under the conditions an application will encounter.
A tool call is not authorization
The documentation permits function calling or web search, but not both together in one configuration. This is a concrete workflow choice for developers. Supported function calling should not be confused with confirmed execution of a person's intent. The model's ability to request an operation and the application's decision to carry it out remain distinct concerns.
In a hypothetical customer service assistant, a separate step could require approval before changing an address or submitting a request. That is our proposed architectural boundary, rather than a claimed built-in model feature. Before deployment, a team would compare the suitable region, history handling, protocol and response to ambiguous speech. This evaluates the new voice capability against a completed task, without treating a pleasant conversation as evidence of reliable automated action.
*AI illustration: a conceptual editorial image, not a documentary photograph or an actual product screen.*
*Editorial verification: 2026-10-28T07:29:45.233660+00:00.*
