Qwen3.5-Omni

The Qwen3.5 generation of the Qwen-Omni line: an end-to-end any-to-text-and-speech model that, unlike its open-weight predecessors, ships API-only on Alibaba Cloud Model Studio as qwen3.5-omni-plus and the cheaper qwen3.5-omni-flash. The report says it "scales to hundreds of billions of parameters" without giving a count, uses a Hybrid Attention MoE for both the Thinker and the Talker, takes a new AuT audio transformer for perception, and introduces ARIA, which dynamically aligns text and speech units during streaming synthesis to fix the instability caused by mismatched tokenizer rates. Context is 256K; it accepts up to about 3 hours of audio, an hour of video (400 seconds of 720p at 1 FPS in the report's framing), or up to 2,048 images per request, and speaks in 10 languages with instructable emotion, rate, and volume plus zero-shot voice cloning; speech recognition covers 113 languages and dialects and synthesis 36.

Trained on heterogeneous text-vision pairs and more than 100 million hours of audio-visual content. Qwen claims state of the art across 215 audio and audio-visual understanding, reasoning, and interaction subtasks, beating Gemini 3.1 Pro on key audio tasks and matching it on overall audio-visual understanding, and demonstrates "audio-visual vibe coding" (coding from spoken and on-screen instructions) and controllable audio-visual captioning. Weights are not released; parameter counts and training scale are undisclosed. Artificial Analysis scores the Plus tier 20 on Intelligence Index v4.3 and Flash 12. Filed in September 2026 from the technical report, five months after release.