# Qwen3.5-Omni: Alibaba's Multimodal Model for Generating Code from Screen Videos
Alibaba Cloud's multimodal model Qwen3.5-Omni processes text, images, audio, and video, outputting text and speech in real time. Available in Plus, Flash, and Light versions via Offline API and Realtime API. Context window expanded to 256,000 tokens — allowing analysis of up to 10 hours of audio or 400 seconds of 720p video per request.
Scale and Architectural Improvements
Qwen3.5-Omni significantly outperforms its predecessor Qwen3-Omni. Speech recognition expanded to 113 languages and dialects (vs 19 previously), speech synthesis to 36 (vs 10). Thinker and Talker components use Hybrid-Attention MoE. Pretrained on over 100 million hours of multimodal audio-video data.
ARIA technique (Adaptive Rate Interleave Alignment) dynamically synchronizes text and speech tokens in streaming, minimizing word skips and pronunciation artifacts for numbers.
Benchmark Results
Plus version achieved state-of-the-art on 36 audio and audio-video benchmarks, as well as in speech recognition and translation across dozens of languages. In audio understanding, recognition, translation, and dialogue, the model outperforms Gemini 3.1 Pro. In audio-video tasks — on par with Gemini. Speech generation is more stable than ElevenLabs, GPT-Audio, and Minimax on 20 languages.
Text and visual capabilities preserved at the level of same-size unimodal Qwen3.5 models.
Key metrics comparison:
- Context window: 256K tokens (vs 32K).
- Audio: >10 hours per request.
- Video: ~400 sec 720p.
- ASR languages: 113.
- TTS languages: 36.
- Pretraining data: >100M hours.
New Features for Developers
The model supports semantic interruption: distinguishes user speech from noise. Voice cloning, control of speed, volume, and emotions in TTS available. Integrated WebSearch and FunctionCall for expanded functionality.
Emergent capability Audio-Visual Vibe Coding enables code generation from screen video recordings with audio instructions — without text prompts. This feature emerged without targeted training as a scaling byproduct.
Key Takeaways
- 256K token context handles long video/audio sequences.
- Hybrid-Attention MoE in Thinker/Talker boosts multimodal inference efficiency.
- SOTA on 36 benchmarks, outperforms Gemini 3.1 Pro in audio tasks.
- Emergent video coding: generates working code from screen recordings with speech.
- ARIA eliminates streaming artifacts in real time.
Integration into Applications
For mid/senior developers, Qwen3.5-Omni is ideal for real-time tasks: voice assistants, video analysis, multimodal agents. APIs enable embedding into pipelines with custom functions. TTS stability on 20+ languages simplifies global deployments.
Example scenario: passing coding video with audio — the model extracts logic and writes code. Useful for automating tutorials, reviewing screen sessions, or generating scripts from demos.
— Editorial Team
No comments yet.