Moonshot AI, a Chinese AI startup, released Kimi K2.7-Code on June 12, 2026. The open-weight coding model ships on Hugging Face under a Modified MIT license and is also accessible through the Kimi API and Kimi Code platform. Moonshot designed the model specifically for long-horizon software engineering tasks — planning, multi-file editing, tool calls, and debugging across many sequential steps.
Kimi K2.7-Code uses a Mixture-of-Experts architecture with 1 trillion total parameters and 32 billion active per token, spread across 384 expert blocks. The context window is 256,000 tokens. On Moonshot's internal Kimi Code Bench v2, the model scored 62.0 versus K2.6's 50.9, a 21.8% improvement. On MCP Mark Verified, it scored 81.1, edging out Claude Opus 4.8 (76.4), though GPT-5.5 led at 92.9. Moonshot also reports roughly 30% lower reasoning-token usage compared to K2.6, a cost-relevant claim for agentic runs that can span hundreds of steps.
Moonshot AI closed a $1 billion funding round in 2023 and has been iterating on its Kimi K2 series as part of the intensifying open-weight coding-agent race. Other recent competitors include MiniMax M3, Qwen3-Coder-Next, and Cohere's North Mini Code, all released within the past month. Kimi K2.7-Code adds a MoonViT visual encoder (400M parameters) for image and video input, enabling developers to include screenshots and reproduction recordings in the same prompt. The model's thinking mode is mandatory and cannot be disabled via API.
API pricing is notably lower than closed-source rivals: $0.95 per million input tokens, $4.00 per million output tokens, and $0.19 per million on cache hits, compared to Claude Opus 4.8's $5.00/$25.00 rate. Self-hosting, however, requires significant infrastructure — the model weights occupy roughly 595 GB on disk. All headline benchmark numbers are vendor-reported at launch; independent evaluations are pending.