Qwen3.8-LiveTranslate: API Access, Supported Languages, and Open-Source Status
What languages Qwen3.8-LiveTranslate supports, how to get API access today, and whether an open-source release is planned.
What is Qwen3.8-LiveTranslate?
Qwen3.8-LiveTranslate is a real-time simultaneous interpretation model from the Qwen team, built to translate spoken and written language on the fly while preserving who is speaking and how they sound. The model’s own tagline claims it “names the speaker and carries the meaning,” a nod to its two-part design that tries to keep both the speaker’s identity and the intent of what they said intact across languages.
TL;DR
- Qwen3.8-LiveTranslate supports 60 languages for input, with 29 of them producing full voice output and the rest returned as text only.
- The model is currently API-only, with no confirmed timeline for an open-source or open-weight release.
- Under the hood it uses a hybrid mixture-of-experts architecture split into a “thinker” that streams source and translation together and a “talker” that generates speech in the original speaker’s voice.
- Independent hands-on testing found that short, isolated sentences translate reasonably well, but continuous, real-time conversational flow does not work smoothly out of the box.
- The Qwen team reports benchmark results on faithfulness, fluency, and consistency claiming parity with or an edge over GPT real-time translate and Gemini live translate, along with a notably lower diarization error rate.
- Anyone building with this model today should expect to layer their own pipeline logic around turn-taking and interruption handling rather than relying on the API to manage conversational flow automatically.
What languages does Qwen3.8-LiveTranslate support?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The model accepts input in 60 languages, a spread wide enough to cover most of the globe for a single model. Of those, 29 languages get full voice output, meaning the model both translates the meaning and generates spoken audio in the target language. The remaining languages are limited to text output only, so the translation appears as written text rather than synthesized speech.
This split matters for anyone planning a product around the model. If your use case needs spoken translation output (like a live interpretation booth or a voice assistant), you need to check whether your target language falls into the 29-language voice tier or the text-only tier before building around it. A 31-language gap between input coverage and full voice output is significant, and it’s the kind of detail that can break a use case if it’s assumed rather than checked.
In hands-on testing, languages tried included Russian, Arabic, Dutch, German, Urdu, Hindi, Indonesian, Spanish, Portuguese, and Persian. Results varied: some languages like Indonesian, Spanish, and Persian produced translations described as solid, while others, particularly ones tested with longer or more complex sentence structures, showed more friction.
How does the model’s architecture work?
Qwen3.8-LiveTranslate runs on a hybrid mixture-of-experts setup divided into two functional halves. The “thinker” component streams the source language and its translation simultaneously rather than waiting for a full sentence or turn to finish. The “talker” component handles speech generation, and it’s designed to preserve the original speaker’s voice rather than defaulting to a generic synthetic voice for the translated output.
A key architectural choice is that audio and text are generated in an interleaved stream, not sequentially. That’s meant to reduce the lag between hearing the source language and getting the translated output, which is the whole point of a “live” translation model as opposed to a batch one.
Three additional capabilities are worth noting because they go beyond basic translation:
- The model tries to separate individual speakers and clone each one’s voice separately, useful for multi-person conversations or interviews.
- It can display the source text and its translation side by side, synchronized in time, which helps with review or subtitling workflows.
- It uses prior conversational context to resolve ambiguity, such as linking a nickname mentioned earlier back to a character’s full name mentioned elsewhere in the conversation.
These are meaningful features on paper, and they explain why the model’s marketing leans on words like “meaning” rather than just “translation.” Whether they hold up depends heavily on how conversational the input is, which brings up the access and performance gap.
How do you access Qwen3.8-LiveTranslate right now?
The only current access path is through Qwen’s API. There is no open-weight or open-source release available at this time, and no confirmed date for when or if one is coming. That puts Qwen3.8-LiveTranslate in a different category from many of Qwen’s other releases, which have often shipped as open-weight models that developers can download and self-host.
For teams evaluating the model, this means:
- You need an API key from Qwen to test or integrate it.
- There’s no option today to run it locally or fine-tune it on your own infrastructure.
- Any real-time voice pipeline (handling microphone input, streaming, turn-taking, interruption) has to be built by you around the API, since the API itself does not fully manage conversational flow out of the box.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
That last point came up clearly in independent testing. Running the model through its own provided script, short, isolated sentences translated fine. But once the input resembled a real back-and-forth exchange, the model struggled to detect when a speaker’s turn had ended, and interruption handling did not work as expected. That’s a meaningful gap between the “real-time simultaneous interpretation” framing and what ships by default. Anyone building a production tool on top of this API should plan to write their own turn-detection and streaming logic rather than assuming the API handles it.
Is Qwen3.8-LiveTranslate worth using today?
It depends on what you’re building. If your use case is short, discrete phrases (a translated push-to-talk button, a subtitle generator for pre-segmented clips), the model performs reasonably well across a wide language set. If you’re trying to build a live, flowing conversational interpreter, expect to do real engineering work around the API to handle turn-taking, latency, and interruptions, because that layer isn’t solved out of the box.
The benchmark claims from Qwen position the model competitively against GPT’s real-time translate and Gemini’s live translate offerings on faithfulness, fluency, and consistency, and Qwen reports a diarization error rate (how well the model tells speakers apart) that’s meaningfully lower than both competitors. Those are the vendor’s own reported numbers, and independent testing suggests the practical experience currently lags the marketing on continuous conversational flow, even though isolated-sentence performance is decent.
Given that real-time speech translation is a fast-moving space with GPT and Gemini both fielding their own live translation features, Qwen3.8-LiveTranslate looks like an early, ambitious entry rather than a finished product. The core architecture (streaming thinker/talker split, voice preservation, speaker diarization) is a legitimate advance in design, but the gap between architecture and turnkey usability is still wide.
Frequently Asked Questions
What languages does Qwen3.8-LiveTranslate support for voice output?
It supports 60 languages for input overall, but only 29 of those produce full voice output. The remaining languages return translated text only, without synthesized speech.
Can I download or self-host Qwen3.8-LiveTranslate?
No. As of now, the model is available only through Qwen’s API. There is no open-source or open-weight release, and no announced timeline for one.
Does Qwen3.8-LiveTranslate handle real-time conversation out of the box?
Not fully. Hands-on testing showed that short, standalone sentences translate reasonably well, but the model struggles with detecting when a speaker’s turn ends and with handling interruptions in continuous conversation. Developers need to build additional pipeline logic to manage real-time conversational flow.
How does Qwen3.8-LiveTranslate compare to GPT and Gemini’s live translation features?
Qwen reports that its model matches or beats GPT’s real-time translate and Gemini’s live translate on faithfulness, fluency, and consistency benchmarks, and claims a diarization error rate roughly half that of GPT and a third that of Gemini. These are the vendor’s own reported figures.
What makes Qwen3.8-LiveTranslate’s architecture different from a standard translation model?
It uses a mixture-of-experts design split into a “thinker” that streams the source language and its translation together, and a “talker” that generates speech while preserving the original speaker’s voice. Audio and text are produced in an interleaved stream rather than one after the other, which is meant to reduce translation lag.