Chinese AI company Moonshot AI has open-sourced the weights of its flagship artificial intelligence model, Kimi K3. The company describes it as the world’s first open-weight 3T-class AI model with 2.8 trillion parameters.
The release is considered one of the most significant developments in the AI industry, as leading frontier models from companies such as OpenAI, Anthropic, and Google have largely remained closed-source. Kimi K3 comes with several notable capabilities:
- 2.8 trillion parameters
- A context window of up to 1 million tokens
- Multimodal support for text, images, and video
- Released as an open-weight model
- Optimized for AI agents, software development, and complex reasoning tasks
According to Moonshot AI, Kimi K3 is built on its new Kimi Delta Attention (KDA) and Stable LatentMoE architecture, designed to improve efficiency and performance when handling large-scale workloads.
Moonshot AI claims that Kimi K3 achieves performance comparable to GPT-5.6 Sol and Claude Fable 5 across several benchmark evaluations.
However, these results are based on the company’s own benchmark testing. While early independent reviews also describe the model as highly capable, broader third-party validation is still ongoing.
Although Kimi K3’s weights are publicly available on Hugging Face, running the model locally requires exceptionally powerful hardware. According to community estimates:
- The full model requires approximately 1.4–1.7 TB of memory.
- Even quantized versions may need 650 GB to 1 TB of RAM.
As a result, Kimi K3 is expected to be used primarily by organizations with access to large GPU clusters or cloud AI infrastructure. Kimi K3 is the latest example of Chinese AI companies embracing an open-weight strategy. Following releases from DeepSeek and Z.ai, Moonshot AI is also making a frontier-scale model available to developers.
By providing open model weights, developers and enterprises can fine-tune the model, deploy it on their own infrastructure, and build AI-powered products without relying entirely on proprietary APIs. As competition intensifies, the race is no longer just about building the most powerful AI model, but also about creating the most open and developer-friendly ecosystem.
