Qwen's new 27B model is knocking on Opus 4.6's door. The lab is officially open-sourcing Qwen3.8-27B, and the official benchmarks show this 27B-parameter local model beating Meta's just-released 30B Muse Glimmer in all eight directly comparable tests. The gaps are especially stark on agent and coding benchmarks: Terminal-Bench 2.1 scores 73.0 for Qwen3.8-27B against Muse Glimmer's 51.7, SWE-bench Pro comes in at 61.7 versus 51.2, and OSWorld-Verified...