Moonshot Unveils Kimi K3: A 2.8 Trillion Parameter Open-Weight Model Shattering AI Frontiers

Moonshot has officially launched Kimi K3, a groundbreaking 2.8 trillion parameter open-weight model that is quickly establishing itself as a frontier competitor, if not surpassing, leading proprietary models like GPT-56 Soul and Fable 5. Kimi K3 introduces native vision capabilities and an impressive 1 million token context window, built on Moonshot’s proprietary Kimi Delta attention and attention residuals architecture. Benchmarks like Deep SWE, Frontier SWE, Terminal Bench, Program Bench, and SWE Marathon show Kimi K3 consistently performing at or near the top, often outperforming previous open-weight leaders and even proprietary models in specific tasks. Notably, it leads in Spreadsheet Bench and scores exceptionally high on Browse Comp, hinting at robust agentic and computer-use capabilities. Its AI Omnition score also positions it as one of the most honest open-weight models, exhibiting a reduced tendency to hallucinate.

Beyond benchmarks, practical testing reveals Kimi K3’s strong long-horizon coding performance, adeptness at UI/frontend development, impressive 3D reasoning and game development, and even advanced applications like GPU kernel optimization and chip design—areas where other frontier labs have often imposed restrictions or not focused. This model’s sheer scale, roughly 2.8 terabytes (FP8) or 1.4 terabytes (FP4), necessitates supercomputing infrastructure, with Moonshot recommending 64+ H100 accelerators, making it unsuitable for local deployment but a significant leap in accessible large-scale models. Current usage is limited to Moonshot’s API and subscription platform, with the full weights planned for release on July 27th. While offering competitive pricing (e.g., $15/million tokens out) against proprietary models like Sonnet and even being token-efficient, its origin from a Chinese company raises data privacy and security considerations for sensitive workloads. The open-weight nature of such a powerful model also introduces significant security implications, as it could be leveraged for offensive as well as defensive applications, a concern compounded by Moonshot’s lack of a public system card or detailed safety report at launch.