Kimi K2.5

Kimi K2.5 has strong evidence for coding and agentic developer work. Official Moonshot sources describe an open-source, multimodal, agentic model for coding, visual understanding, documents, research, and multi-step workflows. Public coding evidence is also substantive: SWE-bench Verified lists Kimi K2.5 high reasoning at 70.80% resolved, while Moonshot’s Hugging Face card reports Kimi K2.5 Thinking at 76.8 on SWE-Bench Verified using an internal 5-run framework. LiveCodeBench evidence reports 85.0, ranked 9th among 31 models shown. The model’s broader capability profile is good but not uniformly frontier. Artificial Analysis lists a 47 Intelligence Index, 256k context, 44.2 output tokens/s for the reasoning model, and prices of $0.58/M input and $3.00/M output tokens. Separate provider benchmarking shows much faster non-reasoning hosted options, including Fireworks at 337.8 tokens/s and a lowest blended price of $0.39/M tokens. This supports favorable cost-effectiveness, though availability and performance depend heavily on provider and reasoning mode. Evidence quality is reasonably high because official docs, model pages, GitHub, SWE-bench, LiveCodeBench, Artificial Analysis, Terminal-Bench, and LMArena were checked. Caveats remain: DeepSWE/DataCurve had no Kimi K2.5 row, CodeSOTA only indicates no DeepSWE result, and Terminal-Bench lists “Terminus 2 Kimi K2.5” at 43.2% ± 2.9, which is a weaker agentic terminal result. LMArena ranks are useful public preference signals, not controlled capability measurements.