Breaking

Moonshot AI Open-Sources Coding-Focused Kimi-K2.7-Code

June 12, 2026 at 07:47 EDT

Software Dev & Coding
Foundation Models
Open Source

Moonshot AI released Kimi-K2.7-Code, a coding-focused model, as open source on June 12, 2026. Built on K2.6, it raises several coding scores while cutting reasoning-token consumption by roughly 30%, the company says.

June 12, 2026 · Moonshot AI

Kimi-K2.7-Code: stronger coding, ~30% fewer reasoning tokens

An open-weight coding LLM built on K2.6 — billed as "Kimi's strongest coding model." It curbs overthinking to cut token use while raising agentic and coding scores, with weights published on Hugging Face.

~30%

Fewer reasoning tokens vs K2.6

256K

Context length · multimodal MoE (~1T / ~32B active)

88%

Coding cost saving reported with K2.6 generation

Benchmark gains over K2.6

Percentage-point improvement on Moonshot's own tests

+31.5%

MLS Bench
Lite

+21.8%

Kimi Code
Bench v2

+11.0%

Program
Bench

Pricing — held at parity with K2.6 (CNY per 1M tokens)

¥6.5

Standard input

¥27

Standard output

¥1.3

Cache-hit input

Developers praise

Better instruction-following over long contexts
Greater agent autonomy + strong token efficiency
Improved rate limits / cost after migrating from Claude
Strength in large refactors and visual coding

Caveats

Falls short unless thinking mode is enabled
K2.6 remains stronger for non-coding tasks
Fresh release — large-scale measurements still pending

What's next

"6x High-Speed Mode"

~June 16 · ~180 tok/s, ~260 tok/s short context

→

Open weights

Local inference via vLLM or SGLang

→

3-week charge bonus

API-only campaign running

Continue reading

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