Alibaba's Qwen upgrades its Qwen3-Omni-Flash model
Alibaba's Qwen team has released Qwen3-Omni-Flash-2025-12-01, an upgrade to its native multimodal model that aims to make spoken and audio-visual conversations steadier and gives developers full control over system prompts.
Alibaba's Qwen team announced Qwen3-Omni-Flash-2025-12-01, an updated version of its Qwen3-Omni model. Qwen3-Omni takes text, images, audio and video as input. It produces both text and speech, streaming them at the same time in real time.
The team says the update fixes the "intelligence drop" the model showed in casual spoken conversation and makes multi-turn audio-visual dialogue more stable. Developers can now fully customise system prompts to set persona style, tone and output length. The model handles text in 119 languages, speech recognition in 19 and speech synthesis in 10. Qwen says it has fixed the language-following instability seen in the previous version. It also says speech output now sounds more natural, with better control of speaking rate, pauses and intonation.
Qwen reports benchmark gains over the earlier Qwen3-Omni-Flash:
- Text: ZebraLogic +5.6, LiveCodeBench-v6 +9.3, MultiPL-E +2.7, WritingBench +2.2
- Speech understanding: VoiceBench +3.2, plus a lower word error rate on Fleurs-zh
- Vision: MMMU +4.7, MMMU-Pro +4.8, MathVision_full +2.2
- Video: MLVU +1.6
These numbers are only deltas against the earlier version. The post gives no absolute scores and no comparisons with competing models. It also doesn't say how to access the model, what it costs or whether the weights are open.
Next, Qwen plans to work on multi-speaker speech recognition, video OCR, proactive audio-video learning, and stronger support for agent workflows and function calling.