Kimi K2.7 Code has credible coding capacity, but public evidence is mixed across task types. NVIDIA’s model card reports 74.1 SWE-bench Verified for the native INT4 baseline, while DeepSWE reports 31%±1% pass@1 and 16th of 18 displayed entries. Its Long-Horizon Terminal-Bench result—11th, 0.367 mean reward—also supports a score below the strongest coding anchors. It therefore trails Claude Opus 4.8 (95 coding) and the prior Kimi K2.5 evaluation (86), while materially exceeding Gemma 3’s less agent-oriented 42. Official materials substantiate a 262,144-token context, 1T-parameter/32B-active MoE design, tool calling, JSON/partial modes, OpenAI-compatible API, and image/video input. This makes multimodal I/O comparable to K2.5 but below Gemini 3.5 Flash’s stronger 92. Artificial Analysis lists 47.5 output tokens/s; availability and speed are solid rather than category-leading. Published $0.19 cached-input, $0.95 uncached-input, and $4 output per million tokens are unusually competitive against premium frontier models, with clear official pricing. Public preference is modest: Code Arena places it #27 with 4,677 votes and Agent Arena #22 of 38. LiveCodeBench was checked but has no listed result; the model card’s proprietary coding benchmarks are useful context, not independent confirmation. Open weights, Hugging Face materials, and API compatibility support developer experience, but the divergent SWE-bench, DeepSWE, and terminal results warrant a conservative evidence score.
