Kimi K3 becomes the largest open-weight AI model with AWS deployment support

Moonshot AI has released the open-weight version of Kimi K3, a Mixture of Experts model with 2.8 trillion parameters, making it the largest AI model whose weights are freely downloadable. The company published the model weights on July 27 after first launching Kimi K3 through an API on July 16. The release positions Kimi K3 ahead of other Chinese open-weight models, including DeepSeek V4 Pro, while expanding access for developers and enterprise users.

The model weights are available on Hugging Face in the MXFP4 (Microscaling Floating Point 4-bit) format under a modified MIT license. Kimi K3 is designed with 896 specialized experts, but activates only 16 experts for each processed token. As a result, roughly 104 billion parameters are active during inference, reducing computational demands compared with using the full parameter count.

Kimi K3 also supports a context window of one million tokens and includes native multimodal capabilities for processing both text and images. Moonshot AI says the model introduces two new architectural components. Kimi Delta Attention combines linear and traditional attention techniques to deliver decoding speeds up to 6.3 times faster when handling one-million-token contexts. Stable LatentMoE, the company’s new expert-routing framework, is presented as approximately 2.5 times more scalable than the architecture used in Kimi K2.

Despite the open-weight release, deploying Kimi K3 remains a significant technical challenge. The complete model requires around 1.4 terabytes of storage and high-end GPU infrastructure. To address this, Amazon Web Services published a deployment guide on July 30 describing two implementation approaches built around NVIDIA Blackwell Ultra hardware, specifically the p6-b300 instance type equipped with eight NVIDIA B300 GPUs per node.

Moonshot AI also released two supporting open-source projects alongside the model. MoonEP is an expert-parallel communication library designed to maintain balanced GPU workloads regardless of routing imbalances. AgentEnv provides an execution environment for agent-based AI workflows, extending the ecosystem around Kimi K3 beyond the core model itself.

The launch reflects growing competition among companies developing frontier open-weight AI systems. According to Bloomberg, Moonshot AI founder Yang Zhilin said the company aims to expand its user base by offering greater openness and accessibility than proprietary competitors. The pricing strategy also signals an enterprise focus, with API access costing $3 per million input tokens and $15 per million output tokens, placing Kimi K3 among premium large language models competing at the highest level of the market.