Qwen’s September 18 announcement introduced Qwen3.8-LiveTranslate for real-time simultaneous interpretation. The company says its Interleave architecture connects streaming understanding, text output, and speech generation. Qwen says average lagging, or LAAL, fell from 2.8 seconds to 2.3 seconds. That is a company-reported comparison rather than an independently published benchmark result.
Qwen says Qwen3.8-LiveTranslate supports 60 languages for input audio and output text, while output audio covers 29 languages. The figures do not mean speech can be generated in all 60 languages: the broader count concerns spoken input and text translation. Qwen has not presented the model as free, open-weight, or generally available worldwide in this announcement.
Interleave and the Hybrid-MoE Thinker–Talker design
Qwen describes a two-module, Hybrid-MoE-based Thinker–Talker design. The Thinker combines video, audio, source text, and translation in a temporally interleaved causal sequence. The Talker combines translation with source audio to synthesize speech that Qwen says is intended to preserve the original speaker’s timbre. Qwen calls this a design “connecting streaming understanding, text output, and speech generation through interleaving.”
Qwen describes a causal, temporally interleaved process linking those streams. It provides the architecture and intended outcomes, but not a numerical table that independently quantifies each module’s contribution.
For related Qwen coverage, see Alibaba Adds Qwen3.8-Omni-Flash to Its Multimodal AI Model Line. Qwen3.8-LiveTranslate itself is framed around interpretation and speech generation, and its substantiated claims are limited to the features and evaluations Qwen describes.
Speaker separation returns identifiers and source-language text
Qwen identifies real-time speaker separation, synchronized source-and-translation output, and long-context disambiguation as three capabilities added to the system. For speaker separation, the announcement says the API flow can return speaker identifiers and source-language text alongside a translation. Qwen says these speaker-separation and synchronized-output functions are enabled through session configuration, and that the server can return those fields without a separate ASR interface.
This matters because Qwen’s stated API behavior preserves more than the translated sentence. The return of a speaker ID associates output with a particular speaker, while source-language text offers a parallel record next to the translation. The synchronized presentation feature similarly keeps source and translated material together. Those are product-interface claims by Qwen; the announcement does not provide a customer deployment, adoption total, or partner example demonstrating how organizations are using them.
Long-context disambiguation is the third capability named by Qwen. The company does not assign a numeric context-window figure to the official announcement, so the feature should not be equated with a particular token limit based on this launch post. Its stated role is disambiguation over longer context.
(Related: Qwen3.8-Max Gets a September Upgrade for Coding and Agents)
Qwen’s Omnilingua-MSpeaker and FLEURS evaluation claims
Qwen says it tested Qwen3.8-LiveTranslate on Omnilingua-MSpeaker, which it describes as a multi-speaker long-audio set covering 14 language directions. On that evaluation, Qwen claims the system outperformed mainstream real-time interpretation systems on translation faithfulness, fluency, conciseness, and Diarization Error Rate. The claim is meaningful because it joins translation measures with a speaker-attribution measure, matching the launch emphasis on separation. However, the announcement excerpt does not supply the numerical result table; it does not support a precise margin, a universal ranking, or a claim that the model is superior in every setting.
For a second evaluation, Qwen says it assessed 70 language directions using the public FLEURS audio test set. The company says Qwen3.8-LiveTranslate led the previous generation and current mainstream systems on translation quality, average lagging, speech-recognition accuracy, and speech-synthesis quality. Again, these are Qwen’s stated vendor benchmarks, not independently validated results. Readers can identify the measurement categories the company selected, but should not infer percentages, score differences, or a broader performance hierarchy absent the underlying values.
The company’s latency statement fits this evaluation framing. Qwen says the Interleave architecture reduced LAAL from 2.8 seconds to 2.3 seconds, and separately says its FLEURS evaluation covers average lagging. The announcement does not show whether the 2.8-to-2.3 comparison is the result of a particular FLEURS configuration or another test setting. Preserving that boundary avoids turning a vendor claim into a more expansive conclusion than its published evidence permits.
DataNorth reports a WebSocket endpoint with audio and image inputs
An independent report by DataNorth AI says Alibaba Cloud added Qwen3.8-LiveTranslate-Flash-Realtime to Model Studio. DataNorth reports that the endpoint uses WebSocket, accepts audio plus images, understands 60 languages, and produces translated speech in 29 languages. Its language figures match the division set out in Qwen’s announcement: 60 for understanding or input-audio/output-text coverage, and 29 for speech output.
DataNorth’s report provides implementation-oriented context, but it should be separated from Qwen’s primary launch claims. The official post includes an API example with the endpoint string qwen3.8-livetranslate-flash-realtime, while DataNorth identifies the endpoint as Qwen3.8-LiveTranslate-Flash-Realtime. Neither source in the research record establishes global availability, a universal price, or a customer-specific deployment. DataNorth also reports operational details for a Singapore region, but its page cautions that pricing tables may not expose a stable row across regions; those figures are therefore not a basis for global product terms.
Qwen3.8-LiveTranslate’s documented release case rests on an explicit set of company claims: 2.3-second LAAL versus 2.8 seconds, 60 input-audio/output-text languages, 29 output-audio languages, a two-module Hybrid-MoE design, and streaming features for speaker and source-text handling. DataNorth separately reports the same 60-language understanding and 29-language speech-output figures, as well as WebSocket support with audio and image inputs. The next evidence needed is the numerical evaluation data Qwen has not displayed in the cited announcement and deployment information that goes beyond the endpoint and regional details reported by DataNorth.