Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter model the Chinese startup calls the largest open-source model anyone has released, though bigger closed models exist.
It isn't actually downloadable yet, with open weights due July 27, so for now it only runs through Moonshot's own website, app, and paid access. Bloomberg reported that investors are calling this a 'Kimi moment,'" echoing the shock DeepSeek caused with a similar release last year.
Is the Opus 4.8 comparison real, or just Moonshot talking up its own model? The independent benchmarks say yes. Artificial Analysis, a research firm that tests AI models independently of the companies that build them, gave Kimi K3 a score of 57 on its Intelligence Index, edging out Claude Opus 4.8's 56 and ranking it 4th out of 189 models tracked.
On LMArena, a separate site where users vote on which AI gives the better answer to the same prompt, Kimi K3 debuted in 1st place for coding real websites, beating Claude Fable 5 76% of the time in head-to-head tests. But it's not a clean sweep. Both Moonshot's own numbers and Artificial Analysis's testing still put Kimi K3 behind Claude Fable 5 (59.9) and OpenAI's GPT 5.6 Sol (58.9) on that same Intelligence Index.
Price is where Kimi K3 argues its case best. It costs $3 per million input tokens and $15 per million output tokens, the small chunks of text AI companies bill by. That matches Claude Sonnet 5's price exactly, and comes to about a third of what Fable 5 charges for the same output, though it's the most expensive model any Chinese AI lab has released so far.
Damilare Odedina
Subscribe for free to continue reading this article
Subscribe SubscribeAlready Have an Account? Log In
