Moonshot AI has open-sourced the weights and technical report of Kimi K3, its most capable model. Kimi K3 is a 2.8T MoE model with native visual understanding and a 1M-token context window, offering 2.5x intelligence per unit of compute. The release also includes high-performance attention kernels, MoE communication library, and infrastructure for agent environments.
SynthePulse Insight · AI deep reading
Version 2 · 3 sources
On July 27, 2026, Moonshot AI open-sourced the Kimi K3 model weights and released a technical report. The model is a 2.8T-parameter MoE architecture with native visual understanding and a 1M token context window. The company claims the new architecture delivers a 2.5x improvement in intelligence per unit of computation.
On July 27, 2026, Moonshot AI announced via X the open-sourcing of Kimi K3 model weights and the release of a technical report. The model is described as the "most powerful model": a 2.8T-parameter MoE model with native visual understanding and a 1M token context window.
In addition to model weights, Moonshot AI also open-sourced high-performance attention kernels, MoE communication libraries, and infrastructure for running large-scale agent environments. Hugging Face set up an open-source countdown page for Kimi K3.
Moonshot AI claims the new architecture delivers a 2.5x improvement in intelligence per unit of computation, emphasizing efficiency gains rather than simply increasing parameters. This claim is an official statement and has not been independently verified.
The Verge reported that Kimi K3's performance is said to beat some of the best US company systems at lower cost. Moonshot's plan to release model weights for free, and its explicit targeting of US users, intensifies the debate over whether closed-source US models can maintain dominance against increasingly powerful open-source alternatives.
Open-weight models allow developers to inspect AI capabilities, run on local infrastructure, customize systems, and build new products without relying on a single provider, often at lower cost. However, Kimi K3 is not fully "open source" because key components such as training data, code, model architecture, and configuration methods remain proprietary, and it comes with a restrictive license.
China's support for open-weight models is seen as a combination of practical constraints and political strategy: the open-source ecosystem allows Chinese companies to approach frontier innovation despite limited access to advanced chips and computing power, while aligning with Beijing's industrial strategy to promote widespread adoption of Chinese models, tools, and infrastructure.
This article is based on Moonshot AI's official X account statements, third-party reports (The Verge), and industry observations. All performance claims are official statements and have not been independently verified.
The open-sourcing of Kimi K3 marks an important step in China's AI model open-weight strategy. Its 2.8T MoE architecture and efficiency claims are noteworthy, but actual performance awaits third-party benchmark verification.
Primary report
Primary source
Also covered by 2 sources