Alibaba has released its open-source multimodal AI model Qwen3.8-27B, optimised for local deployment.

According to India’s Officechai on Friday, Alibaba also published the Qwen3.8-27B weights under an Apache 2.0 licence and is moving to compete with single-GPU local models released by rivals such as Meta and Google.

Alibaba will also publish the weights for its top model, Qwen3.8-2.4T-A95B. Developers can now choose between a lightweight model that runs easily on local hardware and a frontier-grade model for major agent tasks.

Qwen3.8-27B is sized to run locally on high-spec MacBook Pro or Mac Studio systems. It processes 262,000 tokens at once by default, and can expand its processing range to up to 1,000,000 tokens when using YaRN technology, which artificially increases the amount of text an AI model can retain.

Qwen3.8-27B is competitive with models several times larger in coding and office-work benchmarks.

In Alibaba’s own benchmarks, performance is on par with Anthropic’s Opus 4.6 Max and it also has vision capabilities.

The most closely watched comparison is Meta’s local model Muse Glimmer-30B, released a few days ago. Meta previously said the model outperforms Qwen3.6-27B in most tests, and Qwen3.8-27B is seen as a response.

In categories that directly compared the two models, Qwen3.8-27B led in all of them. In TerminalBench2.1, it scored 73.0 versus 51.7, a gap of more than 20 points, and it also edged ahead in IFBench and GPQA Diamond.

Unlike its previous approach of open-sourcing small models at launch while first releasing top models in closed form, Alibaba this time published the weights for both models at once.

Benchmark results do not necessarily reflect performance in real-world use, and in agent tasks the harness around a model can be as important as the model itself.

According to a report by The New Stack citing early impressions, there are also comments that the model tends to think for too long. Alibaba’s benchmark results are based on the original checkpoint, and performance could drop in the quantized version that most local users run. AI models store each number very precisely, which makes them large and memory-intensive. Quantization lowers this numeric precision to sharply reduce file size and memory use.