Kimi K3, an open model with 2.8 trillion parameters, designed its own chip

Moonshot AI's Kimi K3 model has developed a microchip in simulation. What its 2.8 trillion parameters change for coding.

Kimi K3, signed by Moonshot AI, reaches 2.8 trillion parameters, which the company claims is the largest open model to date. It features native vision and a one-million-token context window, and is designed for long tasks in coding, reasoning, and knowledge work.

On the architectural side, two building blocks power the model: Kimi Delta Attention, which manages information flow over long sequences, and Attention Residuals, which retrieves representations at different depths of the network rather than stacking them uniformly. The sparsity of the Mixture of Experts goes up a notch, with sixteen experts activated out of eight hundred and ninety-six via a Stable LatentMoE framework. According to Moonshot, the combination improves scaling efficiency by about 2.5 times compared to Kimi K2, KDA speeds up decoding by up to 6.3 times on one-million-token contexts, and AttnRes gains nearly 25% in training efficiency for less than 2% in additional cost.

The company highlights long agentic sessions. In fifteen hours of iterations, the model designed a two-phase kernel algorithm and reduced the round-trip training time from 283.6 ms to 114.4 ms without affecting numerical results. In another trial, a forty-eight-hour autonomous run produced and verified in simulation a chip tailored for a small model built on its own architecture, as a proof of concept.

An external ranking extends these in-house demonstrations. Arena elevates K3 to the top of its Frontend Code Arena, a public leaderboard for frontend code generation: 1679 points and a jump of seventeen spots compared to Kimi K2.6, previously in eighteenth place, with first place in six of the seven evaluated categories, from brand design to data tools, simulations, and content creation tools.

Kimi K3 is running now on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, in maximum thinking mode by default, with the full weights expected to be released as open weights on July 27. Regarding pricing, output is billed at fifteen dollars per million tokens, input at three dollars without cache, and thirty cents with. Moonshot acknowledges that overall performance remains below the most powerful proprietary models, and points out two limitations: a sensitivity to how the reasoning history is fed back to the model, and a tendency to decide on its own when instructions are ambiguous.