Moonshot AI just dropped Kimi K2.7-Code — a coding model built for long, complex software tasks. It boosts performance by 21.8% over its predecessor on Moonshot’s own Kimi Code Bench v2. That’s no small feat in a space crowded with claims and hype.

The model runs on a trillion-parameter Mixture-of-Experts architecture, activating 32 billion parameters per token. This design balances brute force with efficiency. With 384 experts in play, it selects eight per token plus one shared. The context window stretches to a massive 256,000 tokens — ideal for sprawling codebases or multi-file debugging.

K2.7-Code isn’t your average chatbot. It plans, edits, runs tools, and debugs across steps. This isn’t casual conversation. Moonshot targets serious engineering workflows — repo-scale refactors, pull request risk analysis, and continuous integration loops. It even handles multi-modal inputs like images and video, useful for debugging with screenshots or recorded repros.

Token efficiency is the secret sauce. Moonshot claims a 30% cut in “reasoning-token” usage versus K2.6. These tokens represent the model’s internal chain-of-thought, the hidden cost behind every planning step. Cutting that overhead means lower costs, faster responses, and more steps before hitting context limits. For workflows that run hundreds or thousands of steps, that’s a big deal.

Benchmarks show gains across multiple internal suites: +21.8% on Kimi Code Bench v2, +11% on Program Bench, and +31.5% on MLS Bench Lite. K2.7-Code also edges out Anthropic’s Claude Opus 4.8 on tool invocation accuracy, scoring 81.1% versus 76.4% on the MCP Mark Verified benchmark. But raw code generation still favors Claude and GPT-5.5.

Moonshot is playing a two-front game. The model is open-source under a Modified MIT license with weights on Hugging Face, enabling self-hosting for regulated industries. At the same time, it’s accessible via the Kimi API and bundled with Kimi Code, Moonshot’s terminal-first coding agent. Membership plans start at $19 per month, and a “6x High-Speed Mode” is coming June 15, promising output speeds of roughly 180 tokens per second, five to six times faster than the base model.

Pricing undercuts competitors. Kimi K2.7 input tokens cost $0.95 per million, output tokens $4.00 per million, and cache hits just $0.19 per million tokens. By comparison, Anthropic’s Claude Fable 5, launched days earlier, commands prices over 50 times higher per token. Moonshot bets most devs prioritize cost-effectiveness over marginal quality gains. For startups and mid-sized teams, that’s a compelling argument.

The model’s large context window and agentic design suit long-horizon tasks that break simpler models. Moonshot’s swarm-based coordination scales to hundreds of parallel sub-agents running thousands of steps autonomously. This enables complex, multi-module refactors and code reviews without human intervention. It’s the kind of grunt work that eats engineering hours and demands reliable automation.

K2.7-Code is no laptop toy. The Hugging Face weights require about 595 GB of disk space and server-class hardware. Self-hosting demands multiple top-tier GPUs, but API access lets teams sidestep that hurdle. The model also ships with native INT4 quantization to reduce memory footprint.

Moonshot’s move signals a growing trend: open-source coding models that don’t just chase raw benchmark scores but optimize for real-world developer workflows and budgets. Kimi K2.7-Code isn’t the absolute fastest or smartest model, but it offers a rare combination of performance, token efficiency, accessibility, and price. That’s a winning hand for many teams.