Alibaba launches Qwen-Audio 3.1 Realtime Plus with 262K context for real-time voice agents

Alibaba Cloud has released Qwen-Audio 3.1 Realtime Plus as a new end-to-end speech-to-speech model for real-time voice assistants, customer-service systems and conversational agents.

The model was added to Alibaba Cloud Model Studio on September 20, 2026 and expands the Qwen-Audio realtime line with a 262,144-token context window, full-duplex voice interaction, function calling, web search and voice cloning. It accepts audio and text input and produces streaming speech and text responses over persistent realtime connections.

Alibaba says the new model uses the same integration protocol as Qwen-Audio 3.0 Realtime Plus, reducing migration work for applications already built around the previous generation. It retains the earlier voice set while adding eight new system voices, including English-language voices and new Chinese options.

The 262K context window is the most visible architectural change for agent applications. Realtime voice systems continuously accumulate transcripts, instructions, tool results and conversational state; increasing available context gives developers substantially more room before older information must be removed, summarized or moved to external memory.

The release also moves Qwen-Audio further away from being only a conversational speech model. With tool execution, web retrieval and persistent conversational state, it can act as the voice interface for software that performs actions rather than merely answering questions.

··········

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣

QWEN-AUDIO 3.1 COMBINES REAL-TIME SPEECH WITH A 262,144-TOKEN CONTEXT WINDOW.

Alibaba is extending the amount of conversational and agent state that can remain available inside a realtime speech session while preserving the existing integration model.

Qwen-Audio 3.1 Realtime Plus has a context window of 262,144 tokens, or exactly 256 × 1,024 tokens.

Alibaba's release documentation lists the model as a realtime duplex speech system and confirms that the 3.1 generation retains the integration protocol used by 3.0 Plus.

For voice agents, context is consumed by more than the words the user speaks. A production session can accumulate system instructions, text transcripts, audio representations, tool definitions, tool-call arguments, API results, retrieved web information and previous assistant responses.

A larger window therefore supports longer-lived interactions in which the agent must remember what has already happened while continuing to react with low latency.

........

Specification	Qwen-Audio 3.1 Realtime Plus
Release date	September 20, 2026
Model ID	qwen-audio-3.1-realtime-plus
Primary workload	Real-time speech-to-speech conversation
Context window	262,144 tokens
Input	Audio + text
Output	Streaming audio + text
Duplex interaction	Supported
Function calling	Supported
Web search	Supported
Voice cloning	Supported, with regional availability constraints
Integration protocol	Compatible with 3.0 Realtime Plus
New system voices	8 additional voices
Thinking mode	Not supported
Main API	Realtime API / WebSocket

........

Alibaba's realtime architecture is event driven. A client establishes a persistent connection, sends session configuration, continuously streams microphone audio and receives incremental text and audio responses without polling.

For AOQ deployments, audio and control traffic are separated: microphone and generated speech travel over the audio track, while realtime protocol events use a data track. Alibaba's updated implementation guidance for Qwen-Audio 3.1 includes server-side voice-activity detection and low-latency mobile integration.

The practical difference from a large-context text chatbot is that context is being consumed while a latency-sensitive audio stream is active. Developers therefore cannot treat 262K as a target that should routinely be filled. The larger window is better interpreted as a higher ceiling for stateful agents rather than an instruction to retain everything indefinitely.

··········

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣

THE MODEL USES SPEECH-TO-SPEECH INSTEAD OF A THREE-STAGE ASR, LLM AND TTS PIPELINE.

A single realtime model can interpret speech and generate spoken responses without forcing every turn through three separately orchestrated inference systems.

Traditional voice agents commonly use a pipeline: speech → ASR → text → language model → text → TTS → speech.

Each stage can be optimized independently, but every boundary introduces additional processing, serialization and coordination. Alibaba positions Qwen-Audio Realtime as an end-to-end realtime voice-interaction model, allowing speech input and speech output to be handled within one conversational system.

That architecture is particularly relevant when conversation is interruptible. A useful voice assistant has to determine when the user has finished a thought, distinguish a pause from the end of a turn, stop its own speech if the user interrupts and begin processing the next utterance without creating an unnatural delay.

Alibaba exposes multiple turn-control mechanisms across the Qwen-Audio realtime stack. Server VAD detects speech boundaries, smart turn detection combines acoustic signals with semantic interpretation, and developers can also disable automatic turn detection for push-to-talk behavior.

Qwen-Audio can generate both text and speech streams during the same interaction. That makes the model usable in interfaces where audio is presented to the user while text is retained for transcripts, logging, accessibility or downstream software processing.

Alibaba's documented standard realtime audio configuration uses 16 kHz, 16-bit mono PCM for input and 24 kHz, 16-bit mono PCM for output.

The new model does not turn every spoken interaction into an autonomous agent automatically. Agent behavior emerges when the voice model is connected to external tools, application state and permissions. The release is notable because those capabilities now sit directly inside Alibaba's recommended realtime voice model rather than requiring developers to place a separate tool-capable text model behind the speech layer.

··········

[[ADPLUS_300x250_1]]

FUNCTION CALLING AND WEB SEARCH TURN THE VOICE SESSION INTO AN AGENT INTERFACE.

Qwen-Audio 3.1 can decide to retrieve current information or request application functions while maintaining a spoken conversation.

Alibaba lists both function calling and web search as supported capabilities for Qwen-Audio 3.1 Realtime Plus.

Function calling allows an application to expose structured operations to the model. A customer-service agent could query an order database, a travel assistant could inspect reservations, a workplace agent could check a calendar or create a task, and a smart-home interface could translate a spoken request into a device-control operation.

The model determines when a tool is required and produces the structured function request; the surrounding application remains responsible for executing the operation and returning its result.

Web search provides a separate route for information that is newer than the model's internal knowledge. Alibaba's current documentation notes an implementation constraint: web search and function calling cannot be enabled simultaneously within the same configuration, so application architecture must select the appropriate capability or manage the transition externally.

........

Voice-agent component	Qwen-Audio 3.1 capability	Practical role
Spoken interaction	Full-duplex realtime audio	Natural voice conversation
Long session state	262,144-token context	Retains more conversation and tool history
External actions	Function calling	APIs, databases, workflows, device actions
Current information	Web search	Retrieves information beyond static model knowledge
Turn management	Server VAD / smart turn / manual	Determines when the agent should respond
Interruption handling	Streaming realtime architecture	Allows conversational turn changes
Voice identity	System + cloned voices	Branded or personalized assistants
Application migration	3.0-compatible protocol	Reduces integration changes for existing users
Client deployment	WebSocket / realtime transports	Server, mobile and interactive applications

........

The long context becomes more useful when tools enter the conversation. A customer-support call can accumulate identity data, order history, previous troubleshooting attempts, several tool calls and retrieved policy documents.

A 262K-token ceiling gives the agent considerably more room to preserve the original interaction, although production systems should still use explicit memory and retrieval mechanisms rather than relying exclusively on raw context.

The combination also changes the failure surface. A conversational mistake produces an incorrect answer; a tool-enabled conversational mistake can produce an incorrect action. Applications therefore need permission boundaries, argument validation, confirmation flows for consequential operations and reliable logging around tool execution.

··········

EIGHT NEW VOICES EXPAND THE FRONT END, WHILE AVAILABILITY STILL VARIES BY REGION.

The model adds more built-in voice identities and supports cloning, but not every capability is currently documented as equally available across Alibaba Cloud regions.

Alibaba says Qwen-Audio 3.1 Realtime Plus retains the existing system voices and adds eight new ones. Current documentation lists the additional 3.1-compatible voices as longanqian_v3.1, longanhuan_v3.1, longanlingxin_v3.1, longanfengyue_v3.1, xunanchuan, beth_v3.1, betty_v3.1 and cally_v3.1.

Voice selection is configured at the beginning of the realtime session. Alibaba's implementation guidance indicates that the voice setting must be supplied during the first session.update; later attempts to replace it inside the same session are ignored.

The system can also use cloned voices. Alibaba's voice-cloning workflow creates a custom voice from a short reference recording and then supplies the generated voice ID to Qwen-Audio Realtime. The company recommends roughly 10–20 seconds of clean source speech for the relevant workflow.

There is an important regional qualification. Alibaba lists Qwen-Audio 3.1 Realtime Plus for international service scope and China, while dedicated realtime voice-cloning documentation currently limits that specific capability to China (Beijing). The model should therefore not be described as having identical feature availability in every region simply because its model ID is globally listed.

Pricing deserves the same caution. At publication time, Alibaba's public pricing documentation still exposes pricing tables for the previous realtime generation rather than a clearly identified 3.1 Realtime Plus tariff. Using older pricing as though it automatically applies to the new model would create unsupported numerical precision.

Alibaba's current speech-to-speech guidance positions Qwen-Audio 3.1 Realtime Plus for richer voice assistants and customer-service conversations, while keeping the earlier Flash model as the more cost-sensitive alternative.

The significant change is not simply that Alibaba has produced another speech model. Qwen-Audio 3.1 combines realtime duplex audio, a 262K context window, tool execution, web retrieval and configurable voice identity behind a protocol designed for persistent interactive sessions.

For developers building a voice interface that must remember, retrieve information and perform software actions, the model increasingly resembles an agent runtime with speech as its native interface, rather than a conventional voice bot assembled from separate transcription and synthesis services.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

[datastudios.org]