Moonshot AI prepares to release Kimi K3 — the first open model in the 3T class

A page for the upcoming release of Kimi K3 — a new model from the Chinese company Moonshot AI — has gone live on the Hugging Face platform. Unlike many high-profile announcements, this one comes with a countdown to the moment the open weights become available for download right on the same page.

What's promised

Kimi K3 is positioned as the next generation of Moonshot AI's open “frontier” models. The key claim is that this is, in the developers' words, the world's first open model in the 3T class — that is, at the level of three trillion parameters. Its stated area of application is “frontier intelligence” for tasks involving long-horizon programming, knowledge work and reasoning.

Among the technical details, a new architecture is mentioned, built on the Kimi Delta Attention and Attention Residuals mechanisms. Native agentic capabilities are singled out separately: tool calling, browser use and multi-step planning. The extended context window is, by design, meant to let the model keep code on the scale of an entire repository within its field of view.

Open weights and the waiting mode

The main difference from closed competitors is stated outright: the model's weights will be published openly, right here on the release page. Moonshot AI invites anyone interested to hit “Like and Notify” to receive a single email the moment it's published — with no follow-up mailings.

At the time the announcement appeared, more than two thousand users were awaiting the model's release (2,083 by the page's counter), and less than a day remained until launch — the release date is marked as July 27, 2026. The section listing planned artifacts shows just one item so far.

It's worth noting that the available information is limited to a marketing description: the page provides no concrete benchmarks, no comparisons with other models and no training details. The bold claims — from “the first open model in the 3T class” to the quality of the agentic scenarios — can only be verified once the weights are published and the first independent tests come out.

Source: Hacker News