Alibaba's Qwen research lab released Qwen 3.8 27B, a 27 billion parameter, vision capable model under an Apache 2 license, developer Simon Willison wrote on his site. The model supports a 262,144 token context window and runs as a roughly 17 gigabyte quantized file, according to Willison.
Qwen's self-reported benchmarks show the model beating both its predecessor, Qwen 3.6 27B, and the closed weight Qwen 3.7-Plus, one of Qwen's strongest earlier models, Willison reported. He used it successfully for vision tasks including bounding box detection and to generate SVG artwork and small web tools running locally on a MacBook Pro and an Nvidia DGX Spark.
The model's default is "xhigh" reasoning effort, which Willison called "a hilarious default" and "absolutely not a good way to run the model" on consumer hardware. Asked to draw a simple circle, the model burned through extensive reasoning tokens before producing what he called an "absolutely beautiful" but wildly over-engineered result, he wrote. He recommends developers turn the setting down to low or off.
Running locally, the model inferred at only 15 to 30 tokens per second, according to Willison, well below what hosted APIs deliver, which he said limits its use as a daily driver despite its capabilities.
For developers evaluating open weight models at this size, the benchmark numbers are only half the story here. Check the default inference settings before judging speed, because Qwen's own default is tuned for quality demonstrations, not for how anyone would actually run it day to day.