Breaking

Moonshot AI Releases Kimi K2.7 Code, Ranking 5th Among Open Models on Design Arena

June 17, 2026 at 17:47 EDT

Foundation Models
Software Dev & Coding
Open Source

Moonshot AI released Kimi K2.7 Code, a coding-focused model, on June 12, 2026, and open-sourced its weights on Hugging Face under a Modified MIT license. The model ranked 5th among open-weight models (Elo 1312) on the design and code generation benchmark Design Arena as of June 17.

June 12, 2026 · Moonshot AI

Kimi K2.7 Code: An Open-Weight Model Tuned for Long-Horizon Coding

A trillion-parameter Mixture-of-Experts model built on K2.6, posting double-digit benchmark gains while cutting reasoning tokens by roughly 30% — released open-weight under a Modified MIT license.

~1T

Total parameters (MoE), ~32B active per token

256K+

Context window (256K–262K tokens)

~30%

Reduction in reasoning token usage vs K2.6

1312

Design Arena Elo — 5th among open-weight models

Internal benchmarks: K2.6 vs K2.7 Code

Score on each bench — taller column is better. Percentages show the gain.

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%

K2.6 K2.7 Code

Speed

−78.8s

Avg generation time vs K2.6 (337.6s). HighSpeed mode targets 180–260 tokens/s.

API Pricing / M tokens

Input $0.95 · Output $4.00 · Cache hit $0.19

Local run: ~605GB full precision · ~325GB at 2-bit

Praised in practice

Noticeably faster, more efficient reasoning
Handles large, messy codebases
Strong agentic, multi-step tool-call workflows
Compelling cost-performance via API

Cautions raised

Lags GPT-5.5 and Claude Opus 4.8 on complex code
Can trail GLM 5.2 in some cases
Benchmarks lean vendor-internal
Third-party validation (e.g. SWE-Bench) awaited

Where to run it — Day-0 distribution

Hugging Face Ollama Unsloth (GGUF) Cloudflare Workers AI

Continue reading

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