Beijing-based Moonshot AI just dropped the biggest open-weight language model the world has seen. Kimi K3, unveiled on July 16, packs approximately 2.7 trillion parameters into a Mixture-of-Experts architecture, nearly doubling the 1.6 trillion parameters in DeepSeek’s V4 model released back in April.
What Kimi K3 actually is
The model features a context window of 1 million tokens, making it useful for tasks like legal document analysis, codebase review, or synthesizing massive research datasets.
Kimi K3’s API is already live, with the full weight release scheduled for July 27. “Open-weight” means anyone can download and run the model, though the training code and data remain proprietary. It’s an important distinction from fully open-source, but it’s still a massive departure from the walled-garden approach of Anthropic’s Claude or OpenAI’s GPT series.
Moonshot AI has been on a steady trajectory here. Kimi K2 launched in July 2025 with roughly 1 trillion parameters. The K2.6 update followed in mid-2026. And in June, the company released Kimi Work, a desktop agent capable of managing up to 300 parallel sub-agents.
