Breaking

Moonshot AI Releases Coding-Focused Kimi-K2.7-Code as Open-Weight Model

June 12, 2026 at 09:21 EDT

Open Source
Software Dev & Coding
Foundation Models

On June 12, 2026, Moonshot AI released Kimi-K2.7-Code, a large language model specialized for coding, as an open-weight model, making it available via Hugging Face, the Kimi API, and Kimi Code. Built on the previous-generation K2.6, the company says it improves coding and agent performance while cutting reasoning-token usage by roughly 30%.

June 12, 2026 · Moonshot AI

Kimi-K2.7-Code: an open-weight model built to code for hours

Moonshot AI ships an open-source, coding- and agent-tuned successor to K2.6 — published on Hugging Face and live on the Kimi API — with double-digit benchmark gains while cutting reasoning tokens by ~30%.

total MoE parameters (~32B active)

256k

context length, native multimodal

−30%

reasoning tokens consumed at inference

Benchmark gains over K2.6

Score on each coding / agent benchmark — K2.6 vs K2.7-Code

50.9

62.0

Kimi Code Bench v2

+21.8%

48.3

53.6

Program Bench

+11.0%

26.7

35.1

MLS Bench Lite

+31.5%

42.9

46.9

Kimi Claw 24/7

Improved

K2.6 K2.7-Code

What developers welcome

Beats K2.6 at coding using ~30% fewer tokens
Strong long-horizon performance (hours-to-days tasks)
Open weights under Modified MIT terms

Where caution remains

Trillion-class size makes self-hosting hard without serious infra
K2.6 still recommended for general-purpose use
Benchmark selection questioned; long-term reports still limited

The Kimi line so far

K2 (Jul 2025) → K2 Thinking → K2.5 multimodal → K2.6 → K2.7-Code

Continue reading

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