Moonshot AI Releases Kimi K3 Model as Open Source

Chinese AI company Moonshot AI has officially released its Kimi K3 model as open source. The company has made the model's weights, a technical report, and key infrastructure technologies used in its training publicly available.

The Kimi K3 model features 2.8 trillion parameters and is built upon the Kimi Delta Attention (KDA) mechanism. It natively supports visual understanding and offers a substantial context window of one million tokens, allowing it to process vast amounts of information.

Moonshot AI states that Kimi K3 is the first open-source model to reach a parameter scale of 2.8 trillion. The company notes that Kimi models have consistently represented the upper limit in scale for open-source models over the past twelve months.

Technically, Kimi K3 utilizes a Mixture of Experts (MoE) architecture along with the Stable LatentMoE framework, enabling efficient activation among 896 experts. Optimized training methods and data recipes contribute to improved efficiency. Moonshot AI reports that these structural improvements make Kimi K3 approximately 2.5 times more efficient than its predecessor, K2, in terms of scaling.

This model demonstrates strong capabilities in long-range encoding, capable of handling extended engineering tasks and understanding large codebases with minimal human supervision. It also excels at tasks integrating software engineering and visual reasoning, using screenshots and visual feedback to optimize areas like game development and CAD applications. While acknowledging it lags behind top closed-source models like Claude Fable 5 and GPT-5.6 Sol, Moonshot AI asserts Kimi K3 provides cutting-edge performance that surpasses other open-source models in evaluations.