Summary

  • Chinese e-commerce giant Alibaba has released a lightweight version of its Qwen3 multimodal model.
  • Named Qwen2.5-Omni-3B, it can run on consumer-grade hardware, while retaining the functionality of its more complex predecessor, including text, audio, image and video input.
  • According to the manufacturer, it retains 90% of the larger model’s functionality, but uses 50% less GPU memory, meaning it can be used on high-end laptops and desksops, rather than dedicated hardware.
  • However, the licence stipulates that it is for research use only, meaning a separate licence must be obtained from Alibaba to use it in commercial products.
  • The new model is available to download from Hugging Face Transformers, Docker containers or Alibaba’s vLLM.

By Carl Franzen

Original Article