Kimi K2.7 Code

Kimi K2.7 Code, published on 12 June 2026, is a coding-focused branch of Moonshot AI's open line built on top of Kimi K2.6. The body is unchanged — a mixture of experts with one trillion total parameters and 32 billion active per token, 61 layers, 384 experts with eight selected, MLA attention, a 256K context window and the 400-million-parameter MoonViT vision encoder — so what is new is the training, not the shape. Moonshot targets two things at once: end-to-end completion of long-horizon software engineering work, and the cost of getting there. The model uses roughly 30 per cent fewer thinking tokens than K2.6, which matters because reasoning tokens are billed like any other. The vendor's own table pairs it directly against its predecessor and the improvement is consistent: 62.0 against 50.9 on the in-house Kimi Code Bench v2, 53.6 against 48.3 on Program Bench (rebuilding a program's behaviour from a compiled binary and its documentation alone), 35.1 against 26.7 on MLS-Bench Lite, 46.9 against 42.9 on Moonshot's multi-day co-working benchmark, 76.0 against 69.4 on MCP-Atlas and 81.1 against 72.8 on MCPMark-Verified. The same table shows the model still trails GPT-5.5 and Claude Opus 4.8 on most of those tasks, with MCPMark-Verified the one place where it beats Opus. Like K2 Thinking it ships in native INT4, quantisation folded into post-training rather than left to the user, and unlike K2.5 it is sold by its maker: Moonshot's price list carries it at $0.95 per million input tokens on a cache miss, $0.19 on a cache hit and $4.00 for output, with a high-speed variant at double the rate.

▸Videos

No videos yet.

▸Reviews