1. Qwen 3.5 Omni: Architecture, Training, Evaluation, and Omnimodal Capabilities 1.1. Overview of Qwen 3.5 Omni Qwen3.5-Omni is a new-generation fully omnimodal large language model introduced by the Qwen team in 2026. Within a unified framework, it integrates the understanding, reasoning, generation, and action of text, image, audio, and audio-visual content. As a major evolutionary version of the Qwen-Omni model family, Qwen3.5-Omni adopts a Thinker–Talker architecture , in which the Thinker is primarily responsible for text generation and multimodal understanding and reasoning, while the Talker generates streaming speech tokens using the high-level representations provided by the Thinker. The model scales to hundreds of billions of parameters and supports a 256k context length. Architecturally, both the Thinker and the Talker employ a hybrid-attention mixture-of-experts architecture , or hybrid-attention MoE, to improve the efficiency of long-sequence inference. The model undergoes native omnimodal pretraining using large-scale text data, visual data, and more than 100 million hours of audio-visual data, thereby developing strong omnimodal capabilities. For speech generation, the model introduces Adaptive Rate Interleave Alignment , or ARIA, which dynamically coordinates the generation of text units and speech units to improve the stability, naturalness, and prosody of streaming speech synthesis. In addition, the model supports multilingual understanding and speech generation, zero-shot voice customization, controllable audio-visual captioning, and precise audio-visual grounding. Overall, Qwen3.5-Omni is designed as a native omni agent model that can not only perceive and reason across modalities, but also autonomously invoke WebSearch, execute complex FunctionCall operations, generate speech outputs, and conduct real-time streaming interaction, thereby extending omnimodal understanding to tool use and agentic action. ...