Moonshot’s Kimi K3 narrows gap with US models

Moonshot AI’s Kimi K3 has drawn strong interest from scientists after company tests found it can match or outperform rival US models on coding and spreadsheet tasks. The Beijing-based company paused new sign-ups soon after release because demand pushed its systems close to processing limits.

The model is open weight, with core components publicly available for researchers to download and modify, while training details remain private. Moonshot says the weights will be released on 27 July. Researchers cited its scale, 2.8 trillion parameters and a working memory of one million tokens, as a major advance, although that size makes it too large for personal devices and likely requires institutional investment to run.

Scientists said K3 reflects China’s push to build frontier systems and influence global AI governance, arriving near releases from Anthropic and OpenAI and just before the 2026 Artificial Intelligence World Conference in Shanghai. Researchers also noted that Chinese companies are developing competitive models despite US restrictions on advanced AI chips, tightening the gap with proprietary models from Google, Anthropic and OpenAI.