Qwen3-Omni
Alibaba's 30B-A3B omni model that reads text, images, audio and video and answers in text or speech in 10 languages. Apache 2.0.
In short. Qwen3-Omni can listen to speech, watch video, look at images and read text, and it can answer out loud in real time, not only in writing. It activates only about 3B of its 30B parameters per token, but it still needs an 80 GB-class GPU. It is released under Apache 2.0.
Summary
Qwen3-Omni is Alibaba's natively multimodal model that takes text, images, audio and video and answers in text or real-time speech. It uses a Thinker-Talker design. A 30B mixture-of-experts thinker, with about 3B active parameters per token (48 layers, 8 of 128 experts), reads every modality and writes text, and a separate talker turns that into speech by predicting multi-codebook audio codecs and decoding them with a lightweight causal ConvNet. The Instruct checkpoint holds both parts, at 35.3B parameters and 70.5 GB in BF16. The Thinking checkpoint holds the thinker only, with chain-of-thought reasoning and text output, at 31.7B parameters and 63.5 GB. A Captioner fine-tune describes arbitrary audio in detail. The card lists 119 text languages, 19 speech-input languages and 10 speech-output languages: English, Chinese, French, German, Russian, Italian, Spanish, Portuguese, Japanese and Korean.
The technical report, arXiv 2509.17765 of 22 September 2025, says Qwen3-Omni reaches open-source state of the art on 32 of 36 audio and audio-visual benchmarks and overall state of the art on 22, ahead of Gemini 2.5 Pro, Seed-ASR and GPT-4o-Transcribe, and that it matches same-size single-modality Qwen models on text and images. It also gives a theoretical end-to-end first-packet latency of 234 ms in a cold start. These are Alibaba's own results. Alibaba also sells Qwen3-Omni-Flash through its API, and those weights are not published.
Memory demand is high for a 3B-active model. The GitHub README gives theoretical minimums for Transformers in BF16 with FlashAttention 2: for Instruct, 78.85 GB for 15 seconds of video and 144.81 GB for 120 seconds, and for Thinking, 68.74 GB and 131.65 GB. Calling disable_talker() saves about 10 GB when you do not need speech, and the thinker's config sets 65,536 positions. The README says Transformers is very slow on MoE models and recommends vLLM, but vLLM serve currently supports only the thinker, so it returns text and no speech. The license is Apache 2.0, with the text in the GitHub repository, while the Hugging Face repositories carry no license file and tag it as other with the name apache-2.0. In the 30 days to 11 November 2026, Hugging Face counted about 404,000 downloads for Instruct, 261,000 for Thinking and 6,700 for Captioner.
Good for
- Voice assistants that must understand speech, video and images together and answer aloud in one of 10 languages
- Audio understanding beyond transcription, including music, sound effects and mixed audio, with a Captioner checkpoint for detailed audio descriptions
- Speech-to-speech and speech translation work, and agents that call functions from spoken instructions
- Commercial products, since the weights are Apache 2.0
Watch out for
- Plan for an 80 GB-class GPU or several smaller ones, because the card lists BF16 minimums of about 79 to 145 GB for Instruct depending on video length
- Speech output comes only from the Instruct checkpoint, and vLLM serve covers the thinker only, so spoken answers need Transformers, which the README calls very slow for MoE models
- The model dates from September 2025, so compare it with newer omni models such as MiniCPM-o 4.5 and Nemotron 3 Nano Omni on your own audio
- The benchmark claims are Alibaba's, and some columns in the card are for the Flash API models whose weights are not published
- The Hugging Face license metadata and missing license file differ from the Apache 2.0 text on GitHub, so keep the GitHub license with your copy
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Captioner (audio descriptions) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Thinking (thinker only) BF16 | 31.7B 3B active | 73 GB | 36 GB | 69 GB | Workstation |
| Captioner (audio descriptions) BF16 | 31.7B 3B active | 73 GB | 36 GB | 22 GB | Gaming GPU |
| Instruct (thinker and talker) BF16 | 35.3B 3B active | 81 GB | 41 GB | 79 GB | Workstation |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 69 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source multimodal models in 2026 (images, video and audio)
- Open models that listen: nine take audio, but only three answer in speech
Frequently asked questions
- What is Qwen3-Omni?
- Qwen3-Omni can listen to speech, watch video, look at images and read text, and it can answer out loud in real time, not only in writing. It activates only about 3B of its 30B parameters per token, but it still needs an 80 GB-class GPU. It is released under Apache 2.0.
- Can I use Qwen3-Omni commercially?
- Yes. Qwen3-Omni is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Qwen3-Omni open source?
- Qwen3-Omni is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Qwen3-Omni on my own computer?
- Yes, on a good graphics card. The smallest variant, Captioner (audio descriptions), needs about 22 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Qwen3-Omni released?
- Qwen3-Omni was released by Alibaba Qwen on September 22, 2025.
- How long a context window does Qwen3-Omni have?
- Qwen3-Omni supports a context window of up to 64K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Qwen3-Omni-30B-A3B-Instruct model card (Hugging Face) huggingface.co
- Qwen3-Omni repository and README (GitHub) github.com
- Qwen3-Omni Technical Report (arXiv 2509.17765) arxiv.org
- Qwen3-Omni-30B-A3B-Thinking model card (Hugging Face) huggingface.co
- Qwen3-Omni-30B-A3B-Captioner model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| MiniCPM-o 4.5 OpenBMB | 9.37B | 40K | Apache 2.0 (Permissive) | 11 GB Laptop or phone | |
| Nemotron 3 Nano Omni NVIDIA | 31B 3B active | 256K | NVIDIA Open Model (Permissive) | 32 GB Gaming GPU | |
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone |