Moonshot AI releases Kimi K3, largest open-weight model at 2.8 trillion parameters

Moonshot AI released Kimi K3 on July 16, 2025, an open-weight model with 2.8 trillion parameters. The model is the largest openly available model by parameter count to date.

Model specifications

Kimi K3 contains 2.8 trillion parameters, placing it in what the industry categorizes as the 3T class. This significantly exceeds previous open-weight models in size.

The model is designated as "open-weight" rather than "open-source," meaning the weights are available but potentially with restrictions on usage or modification. Specific licensing terms have not been detailed in available reports.

Architecture approach

According to reporting, Moonshot AI's approach with K3 emphasizes memory capacity over raw compute power. The company appears to be making an architectural bet that larger parameter counts and memory access patterns will drive model performance, rather than focusing primarily on training compute or inference optimization.

This represents a divergence from some Western AI labs that have emphasized compute efficiency and smaller, more optimized models in recent releases.

Context and availability

Pricing per token, benchmark scores, context window size, and training data cutoff dates have not been disclosed. Download availability and infrastructure requirements for running the 2.8T parameter model remain unspecified.

Moonshot AI, the Beijing-based company behind the Kimi chatbot platform, has been expanding its presence in the Chinese AI market. The company previously released smaller models in the Kimi series.

What this means

The release of a 2.8T parameter open-weight model signals continued competition in model scale from Chinese AI labs, even as some Western labs have shifted focus to efficiency. The emphasis on memory architecture could indicate different optimization strategies based on available hardware infrastructure in China.

However, parameter count alone does not determine model capability. Without benchmark scores, context window specifications, and performance data, the practical capabilities of K3 relative to smaller models like Meta's Llama 3.1 405B or other open-weight alternatives remain unclear. The accessibility and computational requirements for actually deploying a model of this size may limit its practical adoption outside well-resourced organizations.