Breaking

Moonshot AI Releases Kimi-K2.7-Code, Enters Code Arena

June 13, 2026 at 13:18 EDT

Open Source
Software Dev & Coding
Foundation Models

Moonshot AI released its coding-focused open-source model "Kimi-K2.7-Code" around June 12, 2026, and entered it into "Code Arena: Frontend," an evaluation environment that pits models against each other on web app generation. In tasks that build full web apps and interactive sites from prompts and image uploads, community feedback drives a leaderboard that tests the model's performance.

June 12, 2026 · Moonshot AI

Kimi-K2.7-Code goes open source — frontier coding, fewer thinking tokens

Moonshot AI's most capable coding model yet — a 1T-parameter Mixture-of-Experts built on K2.6, released under a Modified MIT License and pitched as frontier-level performance, cheaply, as open weights.

total parameters (MoE)
~32B active per token

256K

context window range
long-horizon coding

~30%

fewer reasoning tokens
less overthinking

Reported gains over K2.6

Benchmark improvement (%), per Moonshot's own evaluations

+31.5%

MLS Bench
Lite

+21.8%

Kimi Code
Bench v2

+11.0%

Program
Bench

Built for long-running work

Agentic development that chains tool calls and self-corrects across long sessions.

Frontend from prompt + image → Multi-language: Rust · Go · Python → DevOps & performance tuning

Received favorably

Benchmark gains + token efficiency
Substantially cheaper pricing — a possible cost-performance game-changer
Open weights, OpenAI/Anthropic-compatible API

Concerns remain

Gains center on Moonshot's own benchmarks — awaiting third-party verification
KernelBench shows regressions vs K2.6
"More honest, but not more capable"

Continue reading

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