In-depth analysis · About 6 min read
Model Studio lists Qwen Audio 3.1 Realtime Flash in Beijing, with explicit session and tool limits
Alibaba’s October 27 listing and complete model specification separate realtime voice capabilities from context, search, billing and connection limits.
- Published here
- Source published
- Source checked
Alibaba Cloud’s Model Studio update table adds qwen-audio-3.1-realtime-flash to the North China 2 (Beijing) listing on October 27, 2026. The accompanying specification describes a realtime full-duplex text-and-audio model. This establishes a dated platform listing, rather than a verified worldwide model launch or automatic access for every Alibaba account.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The Beijing update table gives October 27 as a calendar date; documentation update clocks do not establish an exact global release time.
- The model specification lists WebSocket, WebRTC and AOQ, voice cloning, Function Calling and web search, with restrictions on combining features.
- Context is listed as 262,144 tokens, with separate input, output and audio-history limits.
- Prices and rate numbers visible in the documentation have unverified regional attribution in this reading; they are not an account quote or a guarantee of unlimited concurrent service.
- This article interprets complete official text and tables; no API call, latency benchmark, account eligibility check or voice-cloning experiment was performed.
What the dated Beijing listing establishes
The relevant news is the model ID’s appearance in the Beijing portion of Model Studio’s model update table. The entry is dated October 27 and attributes realtime text and audio interaction, tool calling, search and voice cloning to this model. A separate model-specific page names Alibaba Cloud Model Studio as the inference provider. Together, the sources identify a particular hosted offering. They do not establish that every Qwen product, consumer subscription or deployment region received the same change.
The update table displays a document update clock of 22:05:53 on October 27; the model page shows 21:00:05 that day. Neither displayed clock specifies its timezone. AISubHub therefore preserves October 27 at day precision for the listing and records the actual reading time separately. It does not convert either documentation clock into a launch timestamp or claim an October 28 product release.
Duplex conversation does not make every feature simultaneous
The specification accepts text and audio as inputs and produces text and audio as outputs. It lists WebSocket, WebRTC and AOQ as access protocols, the default voice longanqian_v3.1, and server_vad, smart_turn and push-to-talk interaction modes. Speaker enhancement is specifically associated with smart_turn. These are the vendor’s documented options; we did not implement a client or establish how well each mode works in a noisy or interrupted conversation.
One restriction matters when designing an assistant: web search is enabled with enable_search and cannot be enabled together with Function Calling. Advertising both capabilities is therefore different from promising both in the same configuration. The page also says enable_thinking is unsupported. Our analysis is that a developer should decide which documented workflow is required before choosing session options. A product cannot assume an unrestricted combination of search, external actions and a separately configurable thinking mode merely because those terms appear across a broader model catalog.
Token context and retained audio have separate ceilings
The page states a 262,144-token context limit, a maximum input length of 245,760 tokens and a maximum output length of 16,384 tokens. These figures identify distinct parts of the documented budget. They are not a promise that a conversation can contain an arbitrary amount of recorded audio, nor a measured guarantee that every maximum can be exercised with identical response speed. The model name’s Flash suffix does not supply a latency benchmark.
Audio history has its own limits: a maximum of 50 turns and 300 seconds of accumulated audio. A voice application consequently needs to consider retained conversational history as well as token accounting. Our practical interpretation is to evaluate what information must survive a long interaction and how the application handles history boundaries. The source does not establish an unlimited audio session, an unlimited recording archive or a specific history-trimming implementation; no such behavior was tested here.
Read the billing rule and connection denominator together
The price table visible in this reading, whose selected region was not established, uses yuan per million tokens and shows text input at CNY 1.5, audio input at CNY 6, text output at CNY 4.5 and audio output at CNY 12. The explanation distinguishes output modes: text-only output is charged at the text-output rate, while combined text-and-audio output charges audio output without an additional charge for its accompanying text. We do not turn these token rates into a per-minute price, because that would require an established audio-to-token conversion and actual usage assumptions.
Beijing and Singapore appear as labels, but the visible table headers and selection state did not establish which region’s prices or rate numbers were shown. These remain quoted documentation values, not a verified quote for one region or account. Regional parity, tax treatment, account eligibility and an actual bill were not checked. The visible rate table lists 60 RPM and 100,000 TPM. Crucially, WebSocket RPM counts connection requests, not model responses. “No separate WebSocket concurrency limit” does not cancel the applicable rate limits or guarantee unlimited throughput. Reconnection behavior and token demand therefore remain application design questions, after the intended region’s terms have been verified.
What developers can evaluate, and what remains unverified
The combination offers a concrete candidate for spoken assistants that need streaming interaction and documented tool or search support. That is an application implication, not evidence that a particular customer-service deployment, language set or voice quality target has already been achieved. The useful evaluation starts with one region, account, protocol and interaction mode, then checks the required capability combination against the documented restrictions and the application’s own needs.
Our evidence covers the complete dated listing entry and the complete visible model-specific specification, including its capability, context, billing and rate tables. The linked realtime guide, integration instructions and client-event reference were not read as supplementary sources. No API request, benchmark, payment, account-entitlement inspection or product demonstration was performed. Global first-release timing, cross-region parity and end-to-end performance remain outside this report.
Why it matters
The listing gives developers a concrete model ID and several deployment constraints to evaluate for realtime voice applications; the feature-combination and session limits may matter as much as the headline context size.
Limits and availability
This is an analysis of Alibaba’s listing and documentation, not a product test. The date is a platform calendar entry, precise release time is unknown, regional parity and account eligibility were not established, and linked integration guides were not used as evidence. Displayed price and rate values were not tied to a confirmed selected region and must not be treated as universally applicable.
Sources and evidence
Source published: · Source checked:
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.