Kimi K3: Open AI Model Rivals US Proprietary Tech

A Beijing-based AI startup, Moonshot AI, claims its new Kimi K3 model has reached and, in some benchmarks, surpassed the level of proprietary AI models from US rivals. Kimi K3 is a 2.8-trillion-parameter model with native vision support and a context window of up to one million tokens. It's described as the first open model in the three-trillion-parameter class. What's interesting is that despite its size, it only uses 16 of its 896 experts for each token, making it more efficient. While Moonshot AI notes Kimi K3 still trails Claude Fable 5 and GPT-5.6 Sol in overall experience, benchmark results tell a different story. Kimi K3 scored 81.2 on FrontierSWE, beating GPT-5.6 Sol and Claude Opus 4.8. It also led on SWE Marathon and set a new record on BrowseComp. The model also performed better on AutomationBench, SpreadsheetBench 2, OmniDocBench, and several visual reasoning tests. Kimi K3 is available through Kimi’s web and desktop apps, Kimi Code, and its API. Moonshot AI plans to release the full model weights by July 27, 2026. This development shows increasing competition in the advanced AI landscape.