.jpg)
Alibaba has launched a 27-billion-parameter open-weight AI model for laptops, days after Meta unveiled its own consumer hardware-focused model, intensifying competition between US and Chinese developers over locally run artificial intelligence.
The Chinese technology group said on Monday that Qwen3.8-27B is designed for consumer hardware and supports coding, professional work, research and long-horizon agentic tasks. Alibaba said the model delivers performance comparable with another model ten times its size.
Alongside the laptop-focused release, Alibaba made the weights of Qwen3.8-Max available for free download, allowing developers to run and build on the model themselves. The company said the model contains 2.4 trillion total parameters, with 95 billion active parameters, although its training data and methods have not been disclosed.
Meta introduced Muse Glimmer last week, a 30-billion-parameter open-weight model designed to run on a Mac or PC using a single consumer graphics card. Meta has said it intends to make the weights of its flagship Muse Spark 1.2 model publicly available as it seeks to strengthen its position against Chinese AI developers and proprietary-model companies including OpenAI and Anthropic.
Data from Hugging Face cited by CNBC showed that Qwen-based models had generated 151,448 derivative works, compared with Meta’s total of about 58,000, giving Alibaba a 2.6-times larger footprint among developers using open-weight models.
Neil Shah, co-founder at Counterpoint Research, told CNBC that the “next battleground” for AI models would be successful deployment on devices rather than relying on data centres. Locally run models can reduce latency and keep data on hardware such as laptops and smartphones, although the computing requirements remain significant.
Alibaba’s model has a 55.6GB unquantised repository, according to The New Stack, while community-created versions for Apple silicon range from about 16.1GB for a four-bit version to 29.5GB for an eight-bit version. The publication said a Mac with 32GB of unified memory could run the lower-precision version at a moderate context length.
The outlet cautioned that benchmark results do not necessarily reflect real-world performance and that quantisation can involve trade-offs in model quality. Alibaba’s reported results nevertheless show substantial gains over its previous Qwen 3.7-Plus model, particularly in coding and knowledge-work tasks.
