Kimi K2.5 tops open-source benchmarks

This is less an “open source wins everything” story than a sign the gap has narrowed enough to matter in production.

• –Claude Opus 4.6 and GPT-5.4 still own the code benchmark, so proprietary models remain the safer bet for pure SWE-heavy workloads.
• –The reasoning result is strong, but not perfectly apples-to-apples: Kimi’s HLE score uses tools, while DeepSeek R1’s 50.2% is pure chain-of-thought.
• –Kimi K2.5 looks like the practical standout because it stays near the top on capability while crushing the runtime metrics that affect user experience.
• –MMLU-Pro is close enough to saturation that 1-2 point gaps matter less than deployment controls, cost, and consistency.
• –The real production advantage of open weights is predictability: versioned behavior, fewer silent model shifts, and more control over upgrade timing.