Moonshot AI Releases Kimi-K3, a 2.8-Trillion Parameter Open-Weight Model
Chinese AI startup Moonshot AI has released Kimi-K3, a massive 2.8-trillion parameter open-weight multimodal reasoning model, on HuggingFace. The model features a 1-million token context window and is designed to rival top models from OpenAI and Anthropic. As one of the largest open-weight models available, Kimi-K3 pushes the boundaries of open-source AI capabilities and provides a benchmark for the economics of serving ultra-large models. Its release intensifies competition among AI providers and could drive down API pricing for high-end reasoning models. Running the model natively in mxfp4 format requires approximately 1.5TB of VRAM, necessitating high-end hardware configurations like 8 to 16 Nvidia B200 GPUs for optimal throughput. On platforms like OpenRouter, the model is priced at $3 per million input tokens and $15 per million output tokens.
## BACKGROUND
Moonshot AI is a Beijing-based artificial intelligence company and one of China's prominent 'AI Tigers.' The company is known for its Kimi series of models, which historically emphasized exceptionally long context windows. Open-weight models allow developers to download and run the model weights locally, though ultra-large models present massive infrastructure and hosting challenges.