Alibaba Unveils Qwen3.5 Omni: A New Multimodal AI Model

The Alibaba Qwen team has introduced Qwen3.5-Omni, a groundbreaking multimodal language model that marks a significant advancement in technology. This model, designed as a competitor to flagship solutions like Gemini 3.1 Pro, offers a unified framework for processing text, images, audio, and video simultaneously within a single computational pipeline.

A key feature of Qwen3.5-Omni is its Thinker-Talker architecture and the use of Hybrid-Attention Mixture of Experts (MoE) across all modalities. This allows the model to handle large context windows and real-time interaction without the traditional latency penalties associated with cascaded systems.

The Qwen3.5-Omni series is available in three variants: Plus, Flash, and Light, which are balanced for performance and cost. The Plus model offers high-complexity reasoning and maximum accuracy, while Flash is optimized for high-throughput and low-latency interaction. Light serves as a more compact option for efficiency-focused tasks.

At the core of the Thinker-Talker architecture are two main components: the Thinker and the Talker. Unlike previous iterations that relied on external pre-trained encoders, Qwen3.5-Omni utilizes a native Audio Transformer (AuT) encoder, pre-trained on over 100 million hours of audio-visual data. This provides the model with a grounded understanding of temporal and acoustic nuances that traditional text-first models lack.

Qwen3.5-Omni has also showcased impressive results on global leaderboards, achieving State-of-the-Art (SOTA) performance on 215 audio and audio-visual understanding tasks. The model surpasses Gemini 3.1 Pro in general audio understanding, recognition, and translation, achieving parity with Google's flagship in audio-visual understanding.

To facilitate real-time interaction, the Alibaba team developed technologies such as ARIA (Adaptive Rate Interleave Alignment), which dynamically aligns text and audio inputs. This enhances the naturalness and robustness of speech synthesis without increasing latency.

One of the unique capabilities of Qwen3.5-Omni is Audio-Visual Vibe Coding, allowing coding tasks to be performed based on audio-visual instructions, opening new horizons for developers and users alike.