Breaking

Moonshot AI Releases Coding-Focused "Kimi-K2.7-Code"

June 12, 2026 at 16:36 EDT

Foundation Models
Software Dev & Coding
Open Source

China's Moonshot AI on June 12, 2026 released its latest coding-focused large language model, "Kimi-K2.7-Code," and made it available as open weights. The model is also accessible on the AI-generated design evaluation platform Design Arena.

June 12, 2026 · Moonshot AI

Kimi-K2.7-Code: Moonshot's strongest coding model yet, open-sourced

A 1T-parameter MoE (~32B active) built on K2.6, released under Apache 2.0 with a 256K context, native multimodality and an always-on thinking mode — tuned for long-horizon coding and agentic tasks.

MoE params (~32B active)

256K

Context length

−30%

Reasoning tokens, less overthinking

Benchmark gains over K2.6

Improvement (%) on each coding & agent benchmark

+21.8%

Kimi Code Bench v2

+11.0%

Program Bench

+31.5%

MLS Bench Lite

Kimi API pricing · per million tokens

Cache Hit

$0.19

Input

$0.95

Output

$4.00

Developers welcome

Strong cost-performance and token efficiency
Higher success on long-horizon coding tasks
Apache 2.0; CLI supports MCP & ACP for IDEs
Seen as a Claude Code alternative, esp. in Asia

The catch

Local execution needs hundreds of GB of memory or VRAM
Impractical on personal PCs
6× High-Speed Mode still only "coming soon"

Available now via Kimi API · Kimi Code CLI · Design Arena — with Tool Calling, JSON Mode, Partial Mode and automatic context caching. Alongside Qwen and GLM, positioned to reshape the economics of open-source coding models.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.