Qwen has officially open-sourced Qwen3.8-27B, and the company's own benchmark results came along for the ride. This local model is just 27B parameters, yet across the eight tests where official comparisons were available, it beats Meta's freshly released 30B model, Muse Glimmer, across the board.
The gap is especially striking on agentic and coding tasks. On Terminal-Bench 2.1, Qwen3.8-27B scores 73.0 versus Muse Glimmer's 51.7; on SWE-bench Pro, it's 61.7 to 51.2; and on OSWorld-Verified, 84.3 to 65.9. The smaller model also leads in general reasoning, document, and vision tests.
Even wilder is the matchup against Claude Opus 4.6. Across the 19 benchmarks where both models have recorded scores, Qwen3.8-27B wins 15. It has already pulled ahead on SWE-bench Pro, LiveCodeBench, OSWorld, AndroidWorld, and several vision tasks, while remaining behind on Terminal-Bench, GPQA, HLE, and NL2Repo.
In short, a 27B model you can run on a personal computer is posting official numbers that reach last generation's closed-source flagship territory.
