[Qwen 27B Challenges Opus 4.6: Muse Glimmer Loses All 8 Tests] According to monitoring by 动察 Beating, Qwen 3.8-27B has been officially open-sourced, and the official benchmark scores have been released. This local model, with only 27B parameters, surpassed Meta's recently released 30B model Muse Glimmer in all 8 directly comparable tests listed by the official source. The gap is particularly evident in Agent and programming tasks. On Terminal-Bench 2.1, Qwen 3.8-27B scored 73.0, compared to Muse Glimmer's 51.7; on SWE-bench Pro, the scores were 61.7 versus 51.2; and on OSWorld-Verified, it was 84.3 versus 65.9. It also led in general reasoning, documentation, and visual tests. Even more striking is the comparison with Claude Opus 4.6. Out of 19 tests where both had scores, Qwen 3.8-27B won 15. It has overtaken in SWE-bench Pro, LiveCodeBench, OSWorld, AndroidWorld, and several visual tasks, while still trailing in Terminal-Bench, GPQA, HLE, and NL2Repo. For a 27B model that can run locally on a personal computer, its official benchmark scores have already reached the level of the previous generation's closed-source flagship. [Original Link]