Qwen’s open-model ecosystem grows with 3 billion downloads and a new release
Alibaba’s Qwen family passed 3 billion downloads, and the team released Qwen 3.8 under Apache 2.0. The downloadable weights give developers another option for adapting the broad open-model family to their own systems.
Artificial Intelligence··Midday
Downloads show the ecosystem's scale
The Qwen family reached 3 billion downloads over the past six months, according to the Hugging Face open-model report relayed by Fortune. The same report counts 418 million downloads for Google and 227 million for Meta. Alibaba offered a broad model family during this period: more than 460 models in the Qwen family were released as open source, and the ecosystem contains more than 300,000 derivative models. These figures provide a distribution measure of which weights developers choose to download for fine-tuning and deployment. A download does not measure how many people use a finished product; repeated downloads of the same weights or their use as the basis for different derivatives are not equivalent to end-user totals. The Hugging Face report is dated August 14 and contains all the figures in this comparison. It reports both Qwen's download volume and the number of derivatives around the family, while providing no separate measure of end-user reach.[1]
Qwen 3.8 arrives with downloadable weights
The Qwen 3.8 family was released by Alibaba's Qwen team under the Apache 2.0 licence. It includes a dense multimodal model with 27 billion parameters and a larger variant named Qwen3.8-2.4T-A95B; the weights can be downloaded from Hugging Face and ModelScope. The model can process text, images, video, diagrams and documents, including videos lasting several hours. Its native context window is 262,000 tokens and can be extended to 1 million tokens with the YaRN method. A thinking mode can be switched on or off for each query and is enabled by default. Alibaba says a hosted version with a window of 1 million tokens will come to Qwen Cloud. The Qwen team also says the model with 27 billion parameters beats the larger Qwen3.7-Plus on coding and office tasks. The announcement names no independent evaluation of that comparison. The statement about coding and office-task performance is presented as the Qwen team's own claim.[2]
Six months of distribution and the August 14 release
Figures reported for the Qwen family over the past six months include 3 billion downloads, more than 460 open-source models and more than 300,000 derivative models. A separate Qwen 3.8 announcement dated August 14 says the new Apache 2.0-licensed weights can be downloaded from Hugging Face and ModelScope. The family includes a dense multimodal model with 27 billion parameters and the Qwen3.8-2.4T-A95B variant. The first figures count downloads during the previous six months and derivatives around the catalogue, while the release information describes the licence, access channels and technical features of specific weights. The model with 27 billion parameters can process text, images, video, diagrams and documents, including videos lasting several hours. Its native context window is 262,000 tokens and can be extended to 1 million tokens with YaRN. Thinking mode is enabled by default and can be changed for each query. A hosted version with a 1 million-token window was announced as a future option for Qwen Cloud. Download totals do not provide an end-user count, and the coding and office-task comparison remains the Qwen team's own claim.[1], [2]
