Moonshot AI unveils Kimi K3, a giant open model rivaling US firms

Moonshot AI unveiled Kimi K3, a 2.8 trillion-parameter open-weight model, as Beijing’s latest push to narrow the gap with top U.S. AI firms. The company said Kimi K3 is the world’s first open 3T-class model and its most capable system yet, aimed at long-horizon coding, knowledge work and reasoning.

Moonshot said the model is built on its Kimi Delta Attention and Attention Residuals architecture, and that it has native vision capabilities plus a 1-million-token context window. Those details matter in a market where model size alone no longer settles the competition, but where access to enough compute and the ability to serve large workloads still determine whether an open model is usable beyond a showcase demo.

The company also drew a line around its own claims. Moonshot said Kimi K3 still trails the most powerful proprietary models, including Claude Fable 5 and GPT 5.6 Sol, even as the launch attracted attention for benchmark comparisons with Anthropic and OpenAI systems. That mix of ambition and caution has fueled scrutiny over whether China’s leading labs are truly closing the gap or only narrowing it in select tasks such as coding and reasoning.

The release lands in a broader technology contest shaped by export controls, restricted compute access and a global scramble to define which AI systems set the standard. For Chinese firms, an open model at Kimi K3’s scale is more than a product milestone. It is a bid to prove that frontier models can be built outside the U.S. leadership circle, even if the hardest part remains turning that model into something most users can actually run.

Moonshot, based in Beijing, has been pressing that case for more than a year. Reuters reported in July 2025 that the company had already released an open-source model as it sought to reclaim market position, and Kimi K3 follows that earlier effort and the Kimi K2.6 release. The new model extends that campaign into the largest open system Moonshot has put forward so far, with the stakes now running from developer adoption to the standards race around future AI infrastructure.