Qwen releases Qwen3.5-Omni omnimodal model family
Alibaba's Qwen team has launched Qwen3.5-Omni, an API-only omnimodal model family that it says beats Gemini-3.1 Pro on most audio benchmarks while matching it on audio-visual understanding.
Qwen announced Qwen3.5-Omni, the successor to Qwen3-Omni. It takes text, image, audio and audio-visual input. The series comes in three Instruct sizes, Plus, Flash and Light, each with a 256k-token context window. According to Qwen, the models can process more than 10 hours of audio or over 400 seconds of 720p video sampled at 1 FPS. Both the "Thinker" and "Talker" components use a Hybrid-Attention mixture-of-experts architecture. Qwen says pretraining was natively omnimodal and included more than 100 million hours of audio-visual data.
Multilingual coverage has grown to speech recognition in 113 languages and dialects and speech generation in 36. Qwen reports that Qwen3.5-Omni-Plus set state-of-the-art results on 215 audio and audio-visual subtasks. In its own comparisons, Plus scores 93.1 on VoiceBench against Gemini-3.1 Pro's 88.9, and 1.11 word error rate on LibriSpeech-clean against Gemini's 3.36. Gemini still leads on WorldSense, VideoMME with audio (89.0 vs 83.7) and the OmniGAIA tool-use benchmark (68.9 vs 57.2). These are vendor-reported numbers.
The offline API adds structured, timestamped audio-visual captioning and what Qwen calls "Audio-Visual Vibe Coding", which means writing code from spoken and visual instructions. The Realtime API supports:
- semantic turn-taking that ignores backchannels and background noise
- native web search and function calling
- control over voice volume, speed and emotion
- voice cloning
Qwen also describes ARIA (Adaptive Rate Interleave Alignment), a method it says makes streaming speech output steadier.
The models are available through Qwen's Offline and Realtime APIs. The post does not say whether open weights will be released or give pricing.