Qwen3.8-27B outperforms Muse Glimmer in 8 tests and narrows the gap with Claude Opus 4.6

According to Beating Monitor, Qwen has officially open-sourced Qwen3.8-27B, along with its official benchmark results. This 27B-parameter local model outperforms Meta’s newly released 30B model, Muse Glimmer, in all eight directly comparable tests listed by the official team. The performance gap is especially pronounced in Agent capabilities and programming tasks: on Terminal-Bench 2.1, Qwen3.8-27B scores 73.0 versus Muse Glimmer’s 51.7; on SWE-bench Pro, it achieves 61.7 compared to 51.2; and on OSWorld-Verified, it reaches 84.3 versus 65.9. It also leads across general reasoning, document, and vision benchmarks. Even more striking is the comparison with Claude Opus 4.6: in the 19 tests where both models have results, Qwen3.8-27B outperforms it in 15. It has surpassed Claude Opus 4.6 in SWE-bench Pro, LiveCodeBench, OSWorld, AndroidWorld, and multiple vision tasks; it still lags behind in Terminal-Bench, GPQA, HLE, and NL2Repo. A 27B model capable of running locally on a personal computer has already reached the performance level of the previous generation’s closed-source flagship models.