Alibaba releases Qwen3-Next, Qwen3-VL and Qwen3-Omni

Alibaba’s Qwen team released three separate model families across September 2025, extending the Qwen3 generation launched that April: Qwen3-Next, a new base architecture; Qwen3-VL, an updated vision-language line; and Qwen3-Omni, an omni-modal model handling text, image, audio and video in one system.

Qwen3-Next centred on an 80-billion-parameter model that activates only around 3 billion parameters per token. It paired a hybrid attention design — mixing linear “Gated DeltaNet” attention with conventional full attention — with a highly sparse mixture-of-experts layer, which Alibaba said let the base model outperform the earlier dense Qwen3-32B while using a fraction of the training compute and giving substantially higher inference throughput at long context lengths. Qwen3-Omni, released under an Apache 2.0 licence, was described as natively end-to-end: a single architecture trained to understand text, images, audio and video and to generate both text and real-time speech, rather than a speech module bolted onto a text model. Qwen3-VL followed with updated instruct and reasoning (“Thinking”) variants of a 235-billion-parameter mixture-of-experts vision-language model.

All three shipped as open weights, continuing Alibaba’s practice of matching or approaching closed frontier releases while giving developers a licence to fine-tune and redistribute. The cluster of releases landed alongside a wider wave of open-weight activity from Chinese labs through 2025 and reinforced Qwen’s position as one of the most widely used open-weight families on platforms such as Hugging Face.