Alibaba's Qwen releases open-weight Qwen 3.8 models under Apache 2.0

Alibaba's Qwen team has released Qwen 3.8, its latest open-weight family, under the Apache 2.0 license on Hugging Face and ModelScope. The centerpiece for most developers is Qwen3.8-27B, a 27-billion-parameter multimodal dense model that reportedly outperforms the larger Qwen3.7-Plus in coding and office tasks, features improved agent capabilities, and natively handles up to 262,000 tokens of context — extensible to roughly one million tokens using the YaRN method.

The more contentious release is Qwen3.8-Max, a 2.4-trillion-parameter model (Qwen3.8-2.4T-A95B) that Alibaba open-sourced on August 13 — its first open-sourced Max-tier flagship, positioned as competitive with Anthropic's Claude Fable 5. But Alibaba attached mandatory commercial restrictions: large revenue-generating users must share a portion of revenue under the license terms, while most developers can use it free bearing their own compute costs.

That caveat sparked debate. Qwen3.8 topped Hacker News at 893 points, but developers flagged the Max model's text-only, thinking-mode-only open weights and mandatory revenue-sharing as undercutting the 'open' claim. Simon Willison publicly corrected himself ('Argh Qwen 3.8 27B not 3.7!'), underscoring the rapid release churn. Qwen also touted its lead in local inference in Hugging Face's State of Open Models report, posting 'Small models, big real-world impact.'

Competitively, Qwen 3.8 lands in the same crowded week as Muse Glimmer, GLM-5.2, and DeepSeek V4-Pro, reinforcing that Chinese labs now set the pace on open weights. The revenue-sharing terms represent a novel middle path between fully open and closed — watch whether other labs copy it, and whether the 27B model becomes the new default local workhorse over Gemma and Llama derivatives.