Alibaba Qwen 3.8 27B Impresses Despite Overthinking

Alibaba's Qwen research lab has released Qwen 3.8 27B, an Apache 2 licensed, 27-billion-parameter vision-capable large language model. Designed to run locally on consumer hardware, the model is a significant step up from its predecessor, Qwen 3.6 27B. Qwen's internal benchmarks indicate that this new release even outperforms Qwen 3.7-Plus, a powerful closed-weight model from earlier this year. At a quantized size of 17GB, it represents a highly capable option for developers running AI locally on laptops or workstations.

However, early testers have noted a major quirk: the model defaults to an extra-high reasoning effort setting. This default setting can cause the model to spend massive amounts of time and compute on simple tasks. In one test, generating an SVG image of a pelican took 21 minutes, consuming 22,276 reasoning tokens to yield just 3,223 output tokens. When the reasoning effort was turned off entirely, the same prompt took just over two minutes. Users are advised to adjust the reasoning effort to low or turn it off for everyday tasks, and to expand the context limit from the default 8,192 tokens to its maximum of 262,144 tokens to avoid running out of memory.

Despite the overthinking habit, the model demonstrates impressive capabilities in vision and coding. It successfully generated accurate bounding boxes for objects in photos and built a functional HTML interface from a single prompt. It also proved capable of running a local coding agent loop. The main drawback is speed, as it averages 15 to 30 tokens per second on local setups. This is significantly slower than hosted API alternatives like OpenAI 5.6 Sol, which reaches 74 tokens per second, or 5.6 Luna at 184 tokens per second.

To address performance bottlenecks, developers can leverage the model's built-in Multi-Token Prediction architecture. Running the model with speculative drafting in llama.cpp yielded a 72 percent performance boost compared to default setups. This optimization suggests that local hardware can run highly sophisticated, multi-modal models efficiently as serving software continues to mature.