Those coding benchmarks are strong for a 27B, though I always want to see how self-reported numbers hold up on independent runs. Curious whether the Thinking variant's gains on SWE-bench justify its extra token cost in practice.