Qwen3.8-27B Wins 15 of 19 Tests Against Claude Opus 4.6

Open-source Qwen3.8-27B outperforms Muse Glimmer in all 8 benchmarks and beats Claude Opus 4.6 in 15 of 19 tests, highlighting strong local deployment capabilities.

Woofun AI reports that Qwen has officially released Qwen3.8-27B under an open-source license, accompanied by its official performance metrics. This locally deployable model, containing 27 billion parameters, surpassed Meta's newly released 30-billion-parameter Muse Glimmer in all 8 directly comparable tests. The performance gap was most pronounced in agent and programming tasks: on Terminal-Bench 2.1, Qwen3.8-27B scored 73.0 versus 51.7; on SWE-bench Pro, it achieved 61.7 against 51.2; and on OSWorld-Verified, it recorded 84.3 compared to 65.9. It also led in general inference, documentation, and vision-related evaluations.

In comparisons with Claude Opus 4.6, Qwen3.8-27B won 15 of the 19 tests where both models had results. It outperformed Claude Opus 4.6 in SWE-bench Pro, LiveCodeBench, OSWorld, AndroidWorld, and several vision tasks, while trailing in Terminal-Bench, GPQA, HLE, and NL2Repo. This demonstrates that a 27-billion-parameter model capable of running on personal computers has reached performance levels comparable to Meta's previous generation of closed-source flagship models.