AlibabaAudio modelActive
Qwen3.8-LiveTranslate
PopularityRank 652 of 876
1out of 100same as last month
Summary
Qwen3.8-LiveTranslate targets real-time interpretation rather than general text chat. It translates audio and images while producing text and audio, supports 60 languages, and can speak 29 of them. Speaker separation, synchronized bilingual output, and long-context disambiguation matter when a conversation has multiple voices and the attribution matters as much as the words.
The WebSocket Realtime API exposes the endpoint as qwen3.8-livetranslate-flash-realtime and caps input at 49,152 tokens and output at 4,096 tokens within a 53,248-token context. Alibaba exposes the model through Model Studio as an API; its model documentation names no license or weight release. Function calling and structured outputs are outside its contract.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 5 published benchmark results
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.
Qwen vendor-reported LiveTranslate average lagging (LAAL) (Show measurement, test conditions, and source)
- Source value
- 2.3
- Score
- 2.3
- Metric
- average lagging (LAAL) for real-time speech translation
- Unit
- seconds
- Benchmark version
- Qwen vendor report published 2026-09-18
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Qwen
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Qwen's release article says the exact Qwen3.8-LiveTranslate model reduces average lagging from 2.8 seconds for the previous generation to 2.3 seconds. The article separately describes public FLEURS evaluation across 70 language directions but does not identify that corpus as the measurement source for this LAAL number.
- Context
- Vendor-reported latency metric, not an independent evaluation or a translation-quality score. No corpus attribution for this LAAL value, sample count, hardware, endpoint configuration, or uncertainty interval is published alongside it.
Index-Translate Table 8 S2TT, Qwen3.8-LiveTranslate streaming MT judge (Show measurement, test conditions, and source)
- Source value
- 0.833
- Score
- 0.833
- Metric
- Gemini-3.1-Pro per-line translation quality rating, averaged on a 0 / 0.5 / 1 scale
- Unit
- ratio
- Benchmark version
- arXiv:2609.40181v2, Table 8, revision dated 2026-10-12
- Category
- audio
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Index-Translate research paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Third-party paper evaluation of the exact Qwen3.8-LiveTranslate model in streaming mode on an in-house speech-translation set of 140 video windows, 50 to 60 seconds each, with 20 windows per direction across Chinese to English, Japanese, Korean, Spanish, Portuguese, Arabic, and English to Chinese. Gemini-3.1-Pro rated translation lines on a 0, 0.5, or 1 scale. The paper does not report the per-metric scored line count or uncertainty interval.
- Context
- This result is from an in-house evaluation reported by a research group other than the model developer, not Qwen self-report. The model service build or served weight revision is not identified. The 140-window dataset size is not used as a per-metric sampleCount because the paper does not state the valid scored denominator.
Index-Translate Table 8 S2TT, Qwen3.8-LiveTranslate median ASR error (Show measurement, test conditions, and source)
- Source value
- 0.186
- Score
- 0.186
- Metric
- median error rate between source-audio transcription and reference transcript
- Unit
- ratio
- Benchmark version
- arXiv:2609.40181v2, Table 8, revision dated 2026-10-12
- Category
- audio
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Index-Translate research paper
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Third-party paper evaluation of the exact Qwen3.8-LiveTranslate model in streaming mode on an in-house speech-translation set of 140 video windows, 50 to 60 seconds each, with 20 windows per direction across Chinese to English, Japanese, Korean, Spanish, Portuguese, Arabic, and English to Chinese. ASR error is the median rate of source-audio transcription errors against the reference transcript. The paper does not report the per-metric scored line count or uncertainty interval.
- Context
- This result is from an in-house evaluation reported by a research group other than the model developer, not Qwen self-report. The model service build or served weight revision is not identified. The 140-window dataset size is not used as a per-metric sampleCount because the paper does not state the valid scored denominator.
Korean Multi-turn Context Translation Benchmark, Qwen3.8 LiveTranslate full-valid mean MQM penalty (Show measurement, test conditions, and source)
- Source value
- 2.123647604327666
- Score
- 2.123647604327666
- Metric
- Mean raw GEMBA-MQM total penalty across status-valid direct-speech translation cells; lower is better
- Unit
- No unit provided
- Benchmark version
- gemba-mqm-context-v1-milmmt-e4b; unified-15arm-gpt6-luna-audio-20260927-001; dataset SHA-256 765b89dc08b4edba4d29362224bb34bdd9e70856d904fb9f7c4e2ccb623e8879
- Category
- audio
- Direction
- Lower is better
- Sample
- 647
- Source type
- independent-evaluator
- Evaluator
- Korean Multi-turn Context Translation Benchmark, independent experiment report
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. The exact participant route is qwen3.8-livetranslate-flash-realtime (participant qwen38-live-translate-two-voice). The integrated direct-speech run supplies audio-history context plus current-utterance audio and scores Korean items translated into English, Japanese and Simplified Chinese. The frozen data fingerprint is 765b89dc08b4edba4d29362224bb34bdd9e70856d904fb9f7c4e2ccb623e8879; run unified-15arm-gpt6-luna-audio-20260927-001. The benchmark judges status-valid translations with reference-free GEMBA-MQM using google/gemini-3.7-flash:batch through openrouter-batch.
- Context
- The report has 648 expected cells, with 647 valid translated and judged cells for the full score. The model/API serving revision is not published. No confidence interval is reported. The score includes recognition errors and is not an ASR-free translation result.
Korean Multi-turn Context Translation Benchmark, Qwen3.8 LiveTranslate current-utterance CER <=5% conditional mean MQM penalty (Show measurement, test conditions, and source)
- Source value
- 1.3916083916083917
- Score
- 1.3916083916083917
- Metric
- Mean raw GEMBA-MQM total penalty on the participant-specific current-utterance CER <=5% subset; lower is better
- Unit
- No unit provided
- Benchmark version
- gemba-mqm-context-v1-milmmt-e4b; unified-15arm-gpt6-luna-audio-20260927-001; dataset SHA-256 765b89dc08b4edba4d29362224bb34bdd9e70856d904fb9f7c4e2ccb623e8879; conditional current-utterance CER <=5% subset
- Category
- audio
- Direction
- Lower is better
- Sample
- 286
- Source type
- independent-evaluator
- Evaluator
- Korean Multi-turn Context Translation Benchmark, independent experiment report
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. The exact participant route is qwen3.8-livetranslate-flash-realtime (participant qwen38-live-translate-two-voice). The integrated direct-speech run supplies audio-history context plus current-utterance audio and scores Korean items translated into English, Japanese and Simplified Chinese. The frozen data fingerprint is 765b89dc08b4edba4d29362224bb34bdd9e70856d904fb9f7c4e2ccb623e8879; run unified-15arm-gpt6-luna-audio-20260927-001. The benchmark judges status-valid translations with reference-free GEMBA-MQM using google/gemini-3.7-flash:batch through openrouter-batch. This is the participant-specific subset whose recognized current-utterance CER is <=5% inclusive; context-audio CER is not required to be <=5%.
- Context
- This is a model-specific conditional subset, not a shared comparison set and not an ASR-free score. It still includes context/current recognition effects. The 286 judged cells are the valid denominator. The serving revision and confidence interval are not published.
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
0 to 100, higher is more popular
20 of 83 models
pyannote Community-1, pyannote: 73
Nemotron-3-Diarization, NVIDIA: 57
Gemini 3.5 Transcribe, Google: 52
Confucius4-R2T2, NetEase Youdao: 52
NVIDIA Nemotron 3.5 Transcribe, NVIDIA: 52
Granite Speech 5.0 470M TurboCTC, IBM: 52
Kokoro-82M, hexgrad: 51
Nova-3, Deepgram: 51
Fish Audio S2 Pro, Fish Audio: 51
Granite Speech 5.0 470M TurboCTC NC, IBM: 50
Orukeet, Oruk: 49
Lyria 3.5, Google: 47
Gemini 3.5 Live Translate, Google: 46
Phonon-2, Fermion Research: 44
Aura-2, Deepgram: 44
Gemini 3.1 Flash TTS, Google: 44
MiniMax-Music3, MiniMax: 41
Audio8-ASR-Infinite, Edge0 AI: 41
GPT-Realtime 2.1, OpenAI: 41
Qwen3.8-LiveTranslate, Alibaba: 1
Qwen3.8-LiveTranslate
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 6, 2026
Qwen3.8-LiveTranslateMAI-Voice-2.1Eleven v4MAI-Transcribe-2-StreamingCartesia Ink-2
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as Qwen3.8-LiveTranslate, chosen by model line, developer, size, price and release date.
- Qwen3.8-Omni-Flash-RealtimeAlibaba · same developer
- MAI-Transcribe-2-StreamingMicrosoft · released around the same time
- Suno v6Suno · released around the same time
- Suno v6-miniSuno · released around the same time
- Suno v6-wildSuno · released around the same time
- Universal-3.6 Pro RealtimeAssemblyAI · released around the same time
- MAI-Transcribe-2Microsoft · released around the same time
- Nemotron-3-DiarizationNVIDIA · released around the same time
All specifications
Architecture
Input and output
- Capabilities
- Speaker separation, synchronized bilingual output, and long-context disambiguation
- Source: Capabilities
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
API
Further details
Sources and data date
Show all 14 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.