Qwen3-Omni

What it is

Qwen3-Omni is a multimodal model from the Qwen team that handles text, audio and video in a single system: it can read text, listen to speech, watch video and answer either in writing or in synthesized speech. The flagship open release is Qwen3-Omni-30B-A3B, published in September 2025 under the Apache 2.0 licence^[1][2].

How it works

The model uses a mixture-of-experts (MoE) Thinker-Talker design, in which the work of understanding and reasoning is separated from the work of producing speech. Its audio side is built on an audio pre-training stage the team calls AuT, and speech generation uses a multi-codebook scheme intended to keep speech latency low^[1].

That design is aimed at interaction rather than one-shot transcription. The model supports low-latency streaming of audio and video, and is built to handle natural turn-taking in a conversation, so a user can speak, be interrupted, and continue without waiting for a full file to be processed^[1].

Languages and voices

Coverage differs by modality. The model understands text in 119 languages, accepts speech input in 19 languages, and produces speech in 10: English, Chinese, French, German, Russian, Italian, Spanish, Portuguese, Japanese and Korean^[1].

Speech output is available in three voices — Ethan, which is the default, along with Chelsie and Aiden^[1].

Reported results

According to the model card, Qwen3-Omni reached state-of-the-art results on 22 of 36 audio and audio-visual benchmarks, and was the open-source state of the art on 32 of them^[1]. The Qwen team described its speech recognition, audio understanding and voice conversation as comparable to Gemini 2.5 Pro^[1].

On the text and video side, the model card reports 65.0 on AIME25, 86.6 on MMLU-Redux, 69.6 on GPQA and 70.5 on Video-MME^[1]. These are the team's own reported figures rather than independent measurements.

Variants

The release comes in several forms. There is an Instruct model, a text-only Thinking model that produces chain-of-thought reasoning, and a Captioner fine-tuned to write detailed audio captions with few hallucinated details^[1]. The Captioner points at a practical use: turning recordings into descriptions that can be searched, indexed or fed to other systems.

Where it sits in the line

Qwen3-Omni followed earlier Qwen models that were primarily text-based, and extended the family to native audio and video input with speech output. Its open weights under Apache 2.0 made a model of this scope available for outside use and modification^[1][2].

The line continued afterwards with two proprietary successors: Qwen3.5-Omni in April 2026 and Qwen3.8-Omni-Flash on 18 September 2026^[2]. Those later releases were not open in the same way, which makes the September 2025 model a fixed reference point for what the team was willing to publish openly.