Kimi K3, a new artificial intelligence model from Beijing-based Moonshot AI, launched on Thursday, 16 July 2026 as the largest open-weight system ever released, built on a staggering 2.8 trillion parameters.
Parameters are the internal dials a model tunes to make predictions, and 2.8 trillion of them makes Kimi K3 enormous by any measure, as reported by Tom’s Hardware.
Moonshot built it as a mixture-of-experts system, which fires only the parts it needs for a given task.
What makes Kimi K3 different from the frontier models
The launch also lands as a geopolitical marker. China’s AI labs have been working around United States export controls that restrict access to the most advanced chips, and shipping a 2.8 trillion parameter model is a statement that those limits have not stopped progress.
The headline claim is coding. Arena, a blind developer testing platform, ranked Kimi K3 first in its Frontend Code evaluation at 1,679 points, placing it ahead of Anthropic’s Claude Fable 5 in that specific arena of head-to-head matchups judged by real programmers.
The word open-weight matters here. Moonshot plans to release the full model weights on Monday, 27 July 2026, meaning anyone with enough hardware can download and run Kimi K3 themselves, rather than renting access through an application programming interface controlled by the company.
Moonshot is not shy about the achievement.
“K3 stands as Moonshot AI’s most powerful open-source coding model to date,” the company wrote in a press release announcing the launch, framing the model as proof that Chinese labs can build at the very top of the field.
Where Kimi K3 still trails Claude and GPT
Moonshot is honest about the limits. It says Kimi K3 still sits behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall performance, so the crown is narrow rather than absolute.
The win is in front-end coding, not across every benchmark that matters.
For now, Kimi K3 is live through Moonshot’s Kimi app and its Kimi Code developer tools, giving anyone a way to try it before the weights drop.
It also carries a one-million-token context window, enough to hold a small library of documents in a single session.
The pricing tells its own story. Moonshot is charging three dollars per million input tokens and fifteen dollars per million output tokens, with a cheaper cache-hit rate, undercutting the American giants and betting that low cost plus open weights will pull developers toward its stack.
The real test comes on 27 July 2026, when the weights go public and independent researchers can pull Kimi K3 apart on their own machines.
Until then, the benchmark bragging stays a company claim, and the wider AI field waits to see whether it holds up.
