Kimi K2.7 Code HighSpeed review, pricing and limits

Moonshot has not submitted K2.7 Code to any independently audited benchmark suite as of late June 2026. The only numbers on the table are Moonshot's own Kimi Code Bench v2, where K2.7 Code scores 62.0 versus K2.6's 50.9 (a 21.8% relative gain), Program Bench at 53.6 versus 48.3 (+11.0%), and MLS Bench Lite at 35.1 versus 26.7 (+31.5%). On MCP Mark Verified, a third-party tool-invocation benchmark covering Notion, GitHub, Postgres, Filesystem, and Playwright environments, K2.7 Code scores 81.1, ahead of Claude Opus 4.8's 76.4 on that specific test, though Opus 4.8 leads on broader coding suites like Terminal-Bench 2.1. Its closest open-weight rival, GLM-5.2, holds a stronger position on audited benchmarks at 62.1% on SWE-bench Pro, so K2.7 Code's real edge is token efficiency and tool-call reliability rather than raw solve rate.

The defining change from K2.6 is a roughly 30% cut in reasoning-token usage per task, which lowers effective cost on reasoning-heavy coding workloads even though the per-token price is nearly identical. K2.7 Code always runs with extended thinking enabled; a request sent with thinking explicitly disabled is silently rerouted to K2.6 rather than erroring; interleaved thinking is preserved across multi-turn tool-calling sessions for coherent long-horizon coding runs.

The HighSpeed mode is a serving-side optimization, not a retrained model: Moonshot reports substantially faster output on median-length coding completions and bursts even higher on short-context requests, roughly six times the throughput of the standard K2.7 Code deployment. Third-party inference provider Crusoe reports over 430 output tokens/sec running K2.7 on its own optimized stack, which shows the ceiling is provider-dependent rather than fixed by the model.

K2.7 Code HighSpeed is priced per token on Moonshot's native Kimi API, with a cheaper rate available through OpenRouter. The model is also live on Cloudflare Workers AI and the Vercel AI Gateway. Exact rates and self-hosting cost comparisons are broken out below.

Weights are published under a Modified MIT License on Hugging Face and GitHub: free for research and most commercial use, with an attribution clause that kicks in for very large-scale commercial deployments, the same style of condition Moonshot has used since the original K2 release. Self-hosting the FP8 weights needs roughly 1TB of HBM, realistically an 8x H200 SXM5 node, leaving enough headroom for KV cache at full context with small batches. Moonshot also ships native INT4 weights trained with quantization-aware training rather than post-hoc quantization, cutting VRAM roughly in half with minimal quality loss; vLLM, SGLang, and KTransformers are the officially recommended INT4 serving engines.

An independent academic evaluation of the prior K2.5 model found it had a lower overrefusal rate and higher willingness to cooperate with borderline requests than GPT-5.2 and Claude Opus 4.5, a pattern that may carry over until Moonshot publishes its own K2.7 assessment. Training data cutoff is unconfirmed for K2.7 Code specifically; K2.6's was reported at approximately April 2025, though Moonshot never verified that figure itself.