Alibaba open-sources flagship Qwen3.8-Max (2.4T params) as Qwen crosses 3B downloads

Alibaba's Qwen team says its open-weight models have crossed 3 billion cumulative downloads over six months, overtaking Meta (227 million) and Google (418 million) per figures cited in Hugging Face's State of Open Models report. The ecosystem now spans over 460 open-sourced models and more than 300,000 derivatives, and Hugging Face separately noted Qwen leads real-world local inference, followed by Gemma.

The capstone is Qwen3.8-Max: full weights for a 2.4-trillion-parameter model (95B activated per token) with a 1M-token context and native multimodality, released on Hugging Face and ModelScope only ten days after the commercial version launched — the team's first open-sourced Max-tier flagship. Alibaba claims parity with Claude Fable 5 on benchmarks, though it added commercial-use restrictions for the Max weights and launched a $30/year paid tier for its office assistant. A smaller Qwen3.8-27B native multimodal dense model with a 262K context also shipped (Simon Willison publicly corrected initial '3.7' confusion to 3.8-27B).

Strategically, this is the clearest sign yet that Chinese labs are setting the pace in open weights. Apple confirmed it is integrating Qwen into Apple Intelligence for its China market (subject to regulatory approval), and Alibaba Cloud launched a Qwen AI Arena for real-world agent testing plus GB300 deployment guidance co-published with NVIDIA.

The caveat for Western enterprises is licensing: the Max-tier commercial restrictions and China-specific commitments mean 'open' does not equal unrestricted. Still, with GLM-5.3 from Z.ai also rivaling frontier models, the open-model center of gravity has visibly shifted east — the theme running through this week's Hugging Face report.