Beijing-based Moonshot AI released the full model weights for Kimi K3 on July 26, making a 2.8-trillion-parameter system available for free public download, fine-tuning, and self-hosting.
The release makes Kimi K3 the largest open-weight artificial intelligence model distributed to date, and it is currently available on Hugging Face.
Moonshot AI described the system in its technical documentation:
“the world’s first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.”
Architecture and Technical Specifications
Kimi K3 operates on a Mixture-of-Experts architecture containing 2.8 trillion total parameters. The system activates 104 billion parameters per token, selecting 16 out of 896 total experts through its Stable LatentMoE framework.
The model processes up to 1,048,576 tokens in a single context window. It features native multimodal processing for text, images, and video, alongside an always-on reasoning mode.
Moonshot AI built the model using Kimi Delta Attention and Attention Residuals.
The company stated:
“Both architectural updates are designed to help information flow more smoothly through longer sequences and deeper models.”
The model weights ship using MXFP4 quantization with MXFP8 activations. The full download package requires approximately 1.4 terabytes of storage space.
Benchmarks and Performance Testing
On the Artificial Analysis Intelligence Index, Kimi K3 registered a score of 57. That score places the model behind Claude Fable 5 at 60 and GPT-5.6 Sol at 59, but ahead of Claude Opus 4.8 at 56.
Kimi K3 scored 1,679 Elo on the Frontend Code Arena, taking first place in that coding benchmark. AI software firm Cognition confirmed Kimi K3 passed its FrontierCode 1.1 evaluation, adding support for the model into Devin Desktop and CLI tools.
Moonshot AI reported that Kimi K3 performs coding tasks at speeds four times slower than Claude Fable 5, but operates at one-third of the token cost.
Hardware Needs and Cloud Deployment
Running Kimi K3 locally demands specialized hardware infrastructure. A full self-hosted instance requires 64 NVIDIA H100 or B200 GPUs spread across eight servers.
Cloud infrastructure platforms Modal and Together AI enabled day-zero hosting support for Kimi K3. Developers can run serverless endpoints through these services rather than purchasing hardware directly.
Licensing and Resale Rules
Moonshot AI distributed Kimi K3 under a modified license that permits developers to inspect, modify, and fine-tune the model weights. The license allows free self-hosting for enterprise applications and internal tools.
The license introduces restrictions for large-scale cloud providers and commercial inference services. Commercial entities generating more than $20 million in revenue must sign separate agreement terms before reselling Kimi K3 as a managed Model-as-a-Service endpoint.
Moonshot AI designed this rule to prevent third-party cloud vendors from repackaging raw model weights into commercial APIs without revenue agreements.
Other open-weight models from Chinese developers, including Zhipu GLM-5.2 and DeepSeek V4 Pro, continue to use standard MIT licensing without revenue thresholds.
