Moonshot AI today released and open-sourced the Kimi K2.7 Code coding model. 1.1 trillion parameters, a 256K context window, and a fix for a long-standing pain point: the "overthinking" problem in long-horizon coding tasks. Average token consumption is down 30%, and a Speed variant launches next Monday with 5-6x the speed at only 2x the price.

Key Upgrades

Compared with the previous-generation K2.6, K2.7 Code shows clear gains on three dimensions:

Improved coding ability

• Kimi Code Bench v2 up 21.8%
• Program-Bench up 11%
• MLS Bench Lite up 31.5%

Improved agent ability

• Roughly 10% performance gains on the Kimi Claw 24/7 Bench, MCP Atlas, and MCP Mark Verified benchmarks

Improved token efficiency

• Average token consumption down 30%
• Higher performance with fewer tokens

Pricing and Availability

API pricing (same as K2.6):

• Standard input: ¥6.5 per million tokens
• Standard output: ¥27 per million tokens
• Cached input: ¥1.3 per million tokens

How to get it:

• API: via the Kimi API open platform
• Coding plan: the default model on kimi.com/code has been upgraded in sync
• Open source: model weights are available for download on Hugging Face

Note: Thinking mode must be enabled to use K2.7 Code. Turning it off manually causes API errors, and Kimi Code falls back to the K2.6 model. Both the API and Kimi Code enable it by default.

Speed Variant: 6x the Speed, 2x the Price

The K2.7 Code Speed variant launches next Monday (June 15):

• Output speed of about 180 Token/s in typical coding scenarios
• Up to 260 Token/s for short-context scenarios
• Priced at 2x the standard version

Through the Kimi Code Plan early access program, you can try the Speed variant inside Kimi Code, with gradual rollout to Allegretto members and above expected in July.

Limited-Time Promotion

To celebrate the K2.7 Code launch, the Kimi API open platform is running a three-week top-up bonus: top up ¥500 or more and receive a 20%-30% voucher bonus. Vouchers are consumed first when used and are valid for 90 days.

Hands-On Experience

In hands-on testing, the most noticeable impression of K2.7 Code is that it has become "more decisive." The old habit of second-guessing itself on simple tasks and deliberating at length before acting has dropped off sharply.

Test case 1: macOS-style frontend demo

Recreate a macOS-style operating system interface in a single HTML file. K2.7 Code didn't spend much time thinking and jumped straight into development, ultimately delivering a complete boot animation and working basics (Notes, the browser, and other apps all function properly).

Test case 2: agent town

We had K2.7 Code build a recreation of Stanford's "agent town." It first generated a PRD covering product overview, market context, functional architecture, and the technical approach, then developed the MVP under the PRD's guidance. After 30 minutes of continuous iteration it delivered a fully working project with a clean file structure and sensible division of labor.

Quick Start

Coding with K2.7 Code:

• Kimi Code monthly plan: kimi.com/code
• API quick start: platform.kimi.com
• Local deployment: head to Hugging Face to download the model weights

Best use: K2.7 Code is recommended for coding tasks; for non-coding tasks, the more well-rounded K2.6 remains the pick.