Large Language Models

Kimi K2.7 Code

Moonshot's open-weight agentic coding flagship — a 1T-param MoE (32B active) with 256K context and native image/video input, thinking ~30% leaner than K2.6.

Kimi K2.7 Code, released June 12 2026 by Moonshot AI, is a coding-focused refinement of K2.6 rather than a bigger model. It keeps the same ~1-trillion-parameter mixture-of-experts backbone (32B active, 384 experts) and a 256K context (262,144 tokens), with a 400M MoonViT encoder for native image and video input — and it’s released as open weights under a Modified MIT license.

The headline is efficiency: K2.7 Code reorganizes how the model reasons to use roughly 30% fewer thinking tokens than K2.6, which makes long agent loops both cheaper and faster. Moonshot’s own benchmarks show double-digit coding gains (Kimi Code Bench v2 62.0 vs 50.9; MLS Bench Lite 35.1 vs 26.7) and strong tool use (MCP Mark Verified 81.1). At ~$0.95 / $4.00 per 1M tokens — about a fifth of that on cache hits — it lands well under the closed frontier on cost.

The honest caveats: as of mid-June 2026 the numbers are first-party (no independent third-party suites yet), and a 1T-param MoE is non-trivial to self-host, so most teams will reach it through a hosted route. For open, long-horizon coding agents, it’s one of the strongest picks available.

Best for

  • Long-horizon coding agents and repository-scale sessions (256K)
  • MCP / tool-using agentic workflows
  • Vision-to-code from screenshots and PDFs

Pros & cons

Strengths

  • Open weights with a permissive Modified MIT license
  • ~30% fewer reasoning tokens than K2.6 — cheaper, faster agents
  • Strong long-horizon coding and MCP tool use (MCP Mark Verified 81.1)

Limitations

  • Benchmarks are vendor-reported so far
  • A 1T-param MoE is heavy to self-host

Sources

Last updated: 2026-06-18 · Specs and pricing change fast — verify on the vendor's site before relying on them.