Qwen2.5-Omni-3B

Qwen2.5-Omni-3B is the compact sibling of Qwen2.5-Omni, Alibaba's end-to-end omni-modal model that takes text, images, audio, and video as input and answers in text or speech. Released in April 2025 after the 7B flagship shipped on March 26, the 3B variant targets developer GPUs with 24GB of memory and mobile devices.

It inherits the family's Thinker-Talker architecture, which splits perception and speech generation so spoken replies can stream with low latency. Alibaba's own benchmarks put its speech, vision, and text scores close to the 7B model while halving VRAM requirements. Licensing caveat: unlike the 7B's Apache 2.0 release, the 3B's open weights were noted by downstream projects as carrying a research-only, non-commercial license, so confirm licensing before commercial use.