Open models keep closing the gap with frontier coding agents, and Moonshot AI’s latest is the clearest example. Kimi K2.7-Code, shipped June 12, is a 1-trillion-parameter open-weight model built specifically for long-horizon software engineering — planning, editing, running tools, and debugging across many steps rather than one-shot chat.
## What Kimi K2.7-Code does
It’s a Mixture-of-Experts design: 1T total parameters but only 32B active per token, across 384 experts with 8 selected each step, a 256K context window, and multimodal input. The pitch is agentic coding — it’s meant to drive a real software task end to end, and ships with Moonshot’s Kimi Code CLI plus API access. Moonshot reports a 21.8% jump over its predecessor K2.6 on its own Kimi Code Bench v2.
## Open weights and efficiency
The headline for builders is cost. K2.7-Code is open-source under a Modified MIT license with weights on Hugging Face, so you can self-host, and Moonshot says it cuts reasoning token usage by about 30% versus K2.6. API pricing sits around $0.95 per million input tokens. Cheaper tokens plus open weights is exactly how open coding models keep pressuring closed incumbents.
Top AI Product 
Leave a comment