Moonshot AI has unveiled Kimi K3, a 2.8-trillion-parameter open-source AI model that claims frontier-level performance against top proprietary rivals, marking another major step in China’s open AI race.
Moonshot AI has unveiled Kimi K3, claiming it is the world’s largest open-source AI model with 2.8 trillion parameters, surpassing previous Chinese open models including DeepSeek V4 Pro (1.6 trillion parameters) and Zhipu AI’s GLM-5 series (744 billion parameters). The company described K3 as achieving “open frontier intelligence” while targeting long-horizon coding, knowledge work and complex reasoning.
According to Moonshot AI, K3 consistently outperformed OpenAI’s GPT-5.5, Anthropic’s Claude Opus 4.8 and Zhipu AI’s GLM-5.2 across multiple evaluations. On self-reported benchmarks, it also exceeded Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol in Program Bench and SWE Marathon, which assess end-to-end software development and ultra-long-horizon software engineering.
K3 also topped Arena.AI’s Fronted Code Arena across Brand & Marketing, Reference-based Design, Data & Analytics, Consumer Product, and Simulations & Content Creation Tools, trailing Claude Fable 5 only in Gaming.
Independent benchmarking firm Artificial Analysis gave K3 an Intelligence Index score of 57, placing it alongside Claude Opus 4.8 and GPT-5.5, although still behind Claude Fable 5 and GPT-5.6 Sol overall. The firm also noted K3’s higher pricing than most Chinese open-weight peers, with an estimated task cost of US$0.94.
Kimi K3 is available through Moonshot’s chatbot, Kimi Work, Kimi Code and API platform, with full model weights scheduled for release on July 27, reinforcing the company’s open-source strategy amid intensifying AI competition between Chinese and US developers.
