Breaking

Moonshot AI releases coding-focused Kimi-K2.7-Code with open weights

June 12, 2026 at 15:48 EDT

Foundation Models
Software Dev & Coding
Open Source

Moonshot AI released Kimi-K2.7-Code, the latest coding-focused model in its Kimi K2 series, around June 12, 2026, and made the weights available as open source at no cost. Beyond the official Kimi API, it can be accessed via Hugging Face and through third-party gateways such as AI/ML API.

June 12, 2026 · Moonshot AI

Kimi K2.7 Code: an open-weight coding agent that undercuts the closed leaders

Moonshot AI open-sourced its coding-specialized model — a 1T-parameter Mixture-of-Experts with a 256K context — claiming sharp benchmark gains, 30% less reasoning waste, and aggressive pricing aimed at Claude Opus and GPT-5-class systems.

total params · 32B active (MoE)

256K

context window (262,144 tokens)

−30%

reasoning tokens, by curbing overthinking

Benchmark gains vs. K2.6

Company-stated improvement on each suite

+21.8%

Kimi Code
Bench v2

+11.0%

Program
Bench

+31.5%

MLS Bench
Lite

Pricing — per MTok

Cache Hit $0.19

Input $0.95

Output $4.00

Undercuts Claude- and GPT-class systems at open-weight levels.

Closed-loop self-correction

Generate → Fail → Debug → Land ✓

A 2D SpaceX-style flight-physics sim looped through multiple failures and self-corrections before achieving a successful landing.

Praised for

Improved token efficiency
Strength on long-horizon coding tasks
Called a new top open model in coding tools

Caveats

Possible bias in chosen benchmarks
Gaps remain vs. Claude Opus in places
1T MoE is demanding to self-host; long-context weaknesses noted

The trajectory · what's next

K2 → K2.5 → K2.6 → K2.7 Code

A planned 6× High-Speed Mode and broader integration into the Kimi Code CLI and other tooling are expected — positioning open-weight models as serious cost-efficient rivals to closed leaders.

Continue reading

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