Audio modelActive

Qwen3.8-Omni-Flash-Realtime

Alibaba

Released: September 21, 2026
Data date: October 3, 2026

Model class: Omnimodal real-time model
Access: Alibaba Cloud Model Studio API

Qwen3.8-Omni-Flash-Realtime connects live text, audio, and video input to text and audio output. Its realtime documentation lists WebSocket, WebRTC, and AOQ transports; function calling and MCP form the documented action path.

The variant has a 196,608-token input context and up to 65,536 output tokens. Rates differ between Beijing and Singapore. Speech output is billed for both audio and its corresponding text. No separate new weights checkpoint is published.

Specifications and access

Specification	Value and source
Model class	Omnimodal real-time modelSource
Access	Alibaba Cloud Model Studio APISource
Input	Text, audio, and videoSource
Output	Text and audioSource
API model ID	qwen3.8-omni-flash-realtimeSource
Audio history	Up to 100 audio turns or 600 seconds; video up to 50 turns or 240 secondsSource
Transport	WebSocket, WebRTC, or AOQSource
Classification	Real-time API variant without separately published new weightsSource
Singapore rate limit	60 requests and 2M tokens per minuteSource

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

SEED content stability, Chinese (Show measurement, test conditions, and source)

Source value: 0.74
Score: 0.74
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; SEED subsets
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SEED content stability, English (Show measurement, test conditions, and source)

Source value: 0.89
Score: 0.89
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; SEED subsets
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SEED content stability, hard (Show measurement, test conditions, and source)

Source value: 5.29
Score: 5.29
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; SEED subsets
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SpeechSuperClue content stability, multilingual (Show measurement, test conditions, and source)

Source value: 3.34
Score: 3.34
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; multilingual / cross-lingual subsets
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SpeechSuperClue content stability, cross-lingual (Show measurement, test conditions, and source)

Source value: 4
Score: 4
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; multilingual / cross-lingual subsets
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SwanBench-Speech content stability, Chinese (Show measurement, test conditions, and source)

Source value: 1.96
Score: 1.96
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; SwanBench-Speech subsets
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SwanBench-Speech content stability, English (Show measurement, test conditions, and source)

Source value: 2.37
Score: 2.37
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; SwanBench-Speech subsets
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

LongSpeechGeneration content stability, Chinese (Show measurement, test conditions, and source)

Source value: 1.74
Score: 1.74
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; long-horizon generation
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

LongSpeechGeneration content stability, English (Show measurement, test conditions, and source)

Source value: 1.59
Score: 1.59
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 11
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. zero-shot voice cloning; long-horizon generation
Context: Provider-reported zero-shot voice-cloning result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SEED Custom Voice content stability, Chinese (Show measurement, test conditions, and source)

Source value: 0.83
Score: 0.83
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 12
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. custom-voice evaluation; table notes define WER/CER lower, similarity/accuracy/win rate higher
Context: Provider-reported Custom Voice result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SEED Custom Voice content stability, English (Show measurement, test conditions, and source)

Source value: 0.93
Score: 0.93
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 12
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. custom-voice evaluation; table notes define WER/CER lower, similarity/accuracy/win rate higher
Context: Provider-reported Custom Voice result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

SEED Custom Voice content stability, hard (Show measurement, test conditions, and source)

Source value: 5.22
Score: 5.22
Metric: content-stability WER/CER
Unit: %
Benchmark version: arXiv:2609.25611v1, Table 12
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: Qwen3.8-Omni technical report
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. custom-voice evaluation; table notes define WER/CER lower, similarity/accuracy/win rate higher
Context: Provider-reported Custom Voice result; it is not an independently reproduced cohort.

Source Qwen3.8-Omni technical report

Page 1 of 4

Sources and data date

Every statement links to its underlying documentation or leaderboard.

Type	Evidence and data date
Research status	Research date October 4, 2026. 1 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources	independent-verification Original wording No independent reproduction of the exact realtime checkpoint's Table 10-12 results was found in the checked sources.
Additional source	Alibaba Cloud Qwen3.8 Omni Realtime documentation (retrieved October 3, 2026) · Alibaba Cloud Qwen3.8 Omni Realtime documentation · Editorial description reviewed October 3, 2026
Additional source	Alibaba Cloud newly released models (retrieved October 3, 2026) · Alibaba Cloud newly released models · Editorial description reviewed October 3, 2026
Additional source	Alibaba Cloud regional pricing and limits (retrieved October 3, 2026) · Alibaba Cloud regional Realtime pricing · Editorial description reviewed October 3, 2026
Additional source	Qwen3.8-Omni technical report (retrieved October 4, 2026)

Specifications and access

Measurements without matching peer values

Published prices

Sources and data date