Moonshot AI's Kimi K2.7 Code: open-source 1T MoE cuts thinking tokens by 30%

Moonshot AI released Kimi K2.7 Code on Hugging Face, a mixture-of-experts model with 1 trillion total parameters and 32 billion activated per token, using 384 experts with 8 selected at a time. The headline improvement is efficiency: the model uses 30% fewer thinking tokens than its predecessor Kimi K2.6. On Moonshot's own benchmarks, scores improved significantly—Kimi Code Bench v2 jumped from 50.9 to 62.0, Program Bench from 48.3 to 53.6, and MCP Mark Verified from 72.8 to 81.1. However, against closed frontier models like GPT-5.5 and Claude Opus 4.8, K2.7 Code still trails on most metrics, though it edges past Opus on MCP Mark Verified (81.1 vs 76.4). These comparisons are vendor-published, not independently verified.

The model card leaves important questions unanswered. The 'Modified MIT License' modifications are undisclosed, creating uncertainty for product builders. There's no explanation of how the 30% token reduction was achieved or hardware requirements for serving the 1T-parameter MoE—even with native INT4 quantization. Recommended inference frameworks (vLLM, SGLang, KTransformers) put this squarely in the self-hosting-with-real-infra category. The primary beneficiaries are teams that can't route code through external APIs—regulated industries, large infra groups with spare GPUs, and agent-framework developers who value the 30% token cut over marginal leaderboard gains.