Qwen3.5-Omni was released today (March 30, 2026) by Alibaba’s Tongyi Lab. This omnimodal model can understand text, images, audio, and video, and generate both text and speech. Key features: Thinker-Talker architecture with Hybrid-Attention Mixture of Experts, 256K token context, 100M+ hours of multimodal training, 113 language speech recognition, ARIA technology for text-speech alignment, and Audio-Visual Vibe Coding (watch videos and write functional code). Surpasses Gemini 3.1 Pro in audio/video understanding and beats ElevenLabs/GPT-Audio on voice benchmarks. Access via DashScope API or HuggingFace Transformers (80GB VRAM for full model).
Continue reading
Audio-Visual Vibe Coding with Qwen3.5-Omni: Write Code from Video Alone
on SitePoint.
