Qwen3.5-Omni: Scaling Up Toward Native Omni-Modal AGI

Qwen3.5-Omni is the latest fully omnimodal LLM that handles text, image, audio, and audio-visual content all together, offered in three Instruct versions: Plus / Flash / Light. It supports a 256k long-context, and can process over 10 hours of audio input and over 400 seconds of 720P audio-visual input (1 FPS).

Pretraining was conducted in a native omnimodal manner, incorporating over 100 million hours of audio-visual data in addition to text and visual data. As a result, cognition and generation capabilities were strengthened across all modalities, and in particular, multilingual capability improved significantly compared to the previous generation, Qwen3-Omni, now supporting 113 languages/dialects for ASR and 36 languages/dialects for speech generation.

In terms of performance, Qwen3.5-Omni-Plus is reported to achieve SOTA across 215 benchmarks/subtasks covering audio and audio-visual understanding, reasoning, and interaction. It surpasses Gemini-3.1 Pro across overall audio understanding, has risen to the same level in audio-visual understanding, and matches the visual and textual capabilities of the same-sized Qwen3.5 series.

Notably standout features include the following.

• Audio/Audio-Visual Captioning: Generates structured, detailed captions including automatic segmentation, timestamps, and even character relationships
• Audio-Visual Vibe Coding: A new capability that performs coding directly from audio-visual instructions alone
• Realtime API: Supports semantic-based interruption, WebSearch, complex FunctionCall, end-to-end voice control, and voice cloning
• ARIA: Reduces omissions, misreadings, and unstable number pronunciation in streaming speech by dynamically aligning text and speech units

The architecture retains the existing Thinker-Talker structure, but the Thinker receives and processes visual and speech signals via a Vision Encoder and AuT, while the Talker performs context-based speech generation based on the Thinker's multimodal inputs and text outputs. For speech representation, RVQ is used instead of a heavy DiT, and chunk-wise streaming input combined with a streaming Talker design enables real-time interaction.

Compared to the previous generation, the Backbone changed from MoE to Hybrid-MoE, and the sequence length was expanded from 32k to 256k. Speech recognition coverage has also expanded significantly, securing much broader coverage including numerous languages and Chinese dialects.