Breaking

Moonshot AI Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model, Reigniting U.S.-China AI Race

July 20, 2026 at 14:02 EDT

Foundation Models
Open Source

Chinese AI startup Moonshot AI has unveiled Kimi K3, a 2.8-trillion-parameter model it bills as its most capable system yet and, by its own account, the largest open-weight model released to date—an announcement that is prompting researchers to reassess how far China has narrowed the gap with leading U.S. labs.

July 17, 2026 · Moonshot AI

Kimi K3: the largest open-weight model yet — and a new marker in the US–China AI race

Moonshot's 2.8-trillion-parameter Mixture-of-Experts system edges ahead of US frontier labs on coding, trails them on broad benchmarks, and undercuts them on price — with full weights promised days after launch.

2.8T

Total parameters — largest open-weight model to date

16 / 896

Experts activated per token, out of 896 total (MoE)

1.05M

Token context window (1,048,576)

Scaling efficiency vs. Kimi K2

Moonshot claims ~2.5× better scaling than its earlier series

1×

Kimi K2

2.5×

Kimi K3

Coding benchmarks — where K3 edges ahead

Ahead on code, behind on the broad aggregate

Program Bench77.8

SWE Marathon42.0

On coding measures K3 slips past Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol — yet still trails both US leaders across most broader benchmark categories.

Aggressive API pricing (per million tokens)

Input tokens

$15

Output tokens

Plus steep discounts on cached input — paired with a promise to publish full open weights ~10 days after launch.

Developers welcome

Low pricing and open, downloadable weights
Strong long-horizon coding & large-repo work
Native multimodal text, image & video

Cautious observers

Still trails top two labs on overall benchmarks
Heavy hardware needed to run a giant MoE
Weights not out yet — hands-on testing is thin

Continue reading

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