Tutorials, deep dives and product notes — built for developers.
What SWE-bench Pro actually measures, how it works (1,865 tasks, 41 repos, 123 languages), why OpenAI abandoned SWE-bench Verified, the DeepSWE audit that found 32% verifier errors, and how to use coding benchmarks correctly. The definitive explainer.
32B active params vs 10B. $4.00/1M output vs $1.20. 58.6% SWE-bench Pro vs 56.22%. Kimi K2.6 wins on raw performance — but MiniMax M2.7 is the efficiency miracle: 94% of Kimi's coding score at 70% less cost, with only a fraction of the parameters. This is the battle between brute force and architectural genius.
0.2 points apart on SWE-bench Pro. Both open-weight. Both released in April 2026. But the similarities end there. Kimi K2.6 leads on coding (+11.1), agentic tasks (+7.8), and vision. GLM-5.1 counters with pure MIT license, Code Arena #3, and Claude Code compatibility. Here's the definitive comparison.