# Qwen3-ASR Canonical page: https://flit.fyi/glossary/qwen3-asr [Glossary](https://flit.fyi/glossary) / [How it hears you](https://flit.fyi/glossary#recognition) Also: Qwen3-ASR-1.7B, Qwen3-ASR-0.6B Published October 8, 2026 Definition Qwen3-ASR is a family of open-source speech recognition models from Alibaba’s Qwen team, released on 29 January 2026. It recognizes 30 languages and 22 Chinese dialects, and identifies which one is being spoken. It comes in two sizes, Qwen3-ASR-1.7B and Qwen3-ASR-0.6B, with a third model, Qwen3-ForcedAligner-0.6B, for timestamps. All three are released under the Apache 2.0 license. Qwen3-ASR is built differently from older speech models. It is a language model that has been given ears: an audio encoder turns speech into tokens, and a Qwen3 language model writes the text. That is also why it is larger than a model that only transcribes. ## The two sizes The names refer to the language model inside, not to the total. The paper explains that Qwen3-ASR-1.7B is built with Qwen3-1.7B, a projector and an audio encoder of 300 million parameters, and that Qwen3-ASR-0.6B uses Qwen3-0.6B with an encoder of 180 million. So each model is somewhat larger than its name suggests. The larger one is the more accurate. The smaller one is the faster, and is the one to choose when memory is tight. ## Languages and dialects The project’s page states support for language identification and speech recognition in 52 languages and dialects: 30 languages and 22 Chinese dialects, along with English accents from several countries and regions. Language identification is part of the model. You do not have to say which language is coming; it works that out and reports it with the transcript. ## What it can do beyond plain transcription Three features are stated by its makers. - Streaming and offline recognition from the same model, so one set of weights serves both live and recorded speech. - Long audio. The paper gives a limit of 20 minutes for a single piece of speech. - Timestamps, through the separate forced aligner, for up to 5 minutes of speech in 11 languages. ## The limits its makers state The project has no section headed limitations, but its documentation sets some out. Streaming is available with one serving backend only, and streaming does not return timestamps. Timestamps need the second model. And the published speed figures were measured on servers handling many requests at once, so they say little about one person dictating on a laptop. ## Running it on a Mac The original weights are published for Python on a graphics card. For Apple silicon, the same weights are converted to Apple’s MLX format and usually quantized to 4 or 5 bits, which brings the download and the memory use down to laptop size. In Flit Qwen3-ASR is one of the voice models you can choose in Flit. It covers 30 languages, detects yours, and runs on your Mac. [The languages Flit supports](https://flit.fyi/#languages) ## Questions people ask. ### How many languages does Qwen3-ASR support? Its makers state 30 languages and 22 Chinese dialects, 52 in all, with language identification built in. ### Is Qwen3-ASR open source? Yes. The models are released under the Apache 2.0 license. ### What is the difference between Qwen3-ASR-1.7B and 0.6B? The size of the language model inside. The 1.7B model is the more accurate, and the 0.6B model is smaller and faster. ### Does Qwen3-ASR detect the language by itself? Yes. Language identification is part of the model, so the language does not have to be set in advance. ### Does Qwen3-ASR give timestamps? Not by itself. Timestamps come from a separate model, Qwen3-ForcedAligner-0.6B, for up to 5 minutes of speech in 11 languages. ### Can Qwen3-ASR run offline on a Mac? Yes. Converted builds of the published weights run on Apple silicon with no connection once downloaded. ## Related terms - [Language identification](https://flit.fyi/glossary/language-identification) - [Automatic speech recognition](https://flit.fyi/glossary/automatic-speech-recognition) - [MLX](https://flit.fyi/glossary/mlx) - [Quantization](https://flit.fyi/glossary/quantization) - [Streaming transcription](https://flit.fyi/glossary/streaming-transcription) ## Your voice was always faster. 7 days free, then $19 USD once. No account, no subscription, nothing uploaded. [Download for Mac](https://flit.fyi/download) [Buy now · $19 USD](https://flit.fyi/pricing#personal) macOS 14 or later · Apple silicon