Alibaba opens Qwen 3.8-27B: open weights under Apache 2.0

Alibaba's Qwen team has released the Qwen 3.8 model family with open weights under the Apache 2.0 license. At its core is the dense multimodal model Qwen3.8-27B, joined by Qwen3.8-2.4T-A95B, a much larger mixture-of-experts variant for Max-level operation. Both are available on Hugging Face and ModelScope.

The 27B model processes text, images, videos, diagrams and documents and ships with a native context window of 262,000 tokens, extendable to 1 million tokens via the YaRN method. A toggleable "thinking mode" is on by default and can be switched off per query.

According to the Qwen team, the model "outperforms the larger Qwen3.7-Plus in coding and office tasks" and was trained specifically for agent capabilities such as independent planning. Early independent tests praise its vision capabilities but criticize the default reasoning setting, which can consume tens of thousands of tokens on simple tasks.

Memory footprint matters in practice: quantized variants reportedly run in about 17 gigabytes of RAM – bringing a model of this class within reach of well-equipped laptops for the first time, with no cloud connection and no API costs.

For the open-weights market, the release is another rapid-fire move from China: after Meta's Muse Glimmer and DeepSeek's price hikes, Alibaba positions its freely available model as a price-performance attack on Western providers.