Moonshot's Kimi K3 Is the Largest Open Model Yet

Moonshot AI released Kimi K3 on July 16, and the headline number is hard to ignore. At 2.8 trillion parameters, it is the largest open-weight model any lab has put its name to. The Beijing company is not being shy about the comparison it wants you to draw. It says K3 performs competitively with Anthropic's Fable 5, the strongest model in wide release today, and comfortably ahead of Anthropic's Opus 4.8 and OpenAI's GPT-5.6 Sol on its own testing.

A caveat sits underneath that claim. The weights do not go public until July 27, so for now the benchmark results are Moonshot's to report and nobody else's to check. Independent testers get their turn in ten days.

What is actually inside

K3 is a mixture-of-experts model, which means only a fraction of those 2.8 trillion parameters fire on any given token. Moonshot activates 16 of 896 experts per token, so the running cost sits a long way below what the raw parameter count suggests. It takes text, images and video, keeps a reasoning mode switched on by default, and offers a tunable reasoning_effort dial for callers who want to trade speed against depth. The context window is one million tokens.

The engineering note worth flagging is something Moonshot calls Kimi Delta Attention, a hybrid linear attention scheme it credits with up to 6.3 times faster decoding once a context runs into the millions of tokens. Long-context inference is where most models slow to a crawl, so if that figure holds up outside the lab it matters for anyone pushing whole documents or codebases through the model.

On Moonshot's own numbers, K3 leads on coding and agentic tests such as SWE Marathon and BrowseComp while trailing Fable 5 and GPT-5.6 Sol on others. A mixed scorecard, in other words, not a clean sweep. Simon Willison, who ran the model through his usual pelican-on-a-bicycle test on release day, came away impressed enough to say the gap with the Western frontier is now measured in months.

The price is the point

Kimi K3 lists at roughly $3 per million input tokens and $15 per million output, dropping to about $0.30 on cached input. That undercuts most Western flagships, and it continues a pattern we have covered repeatedly this month: Chinese labs shipping open or cheap models that American companies are quietly adopting to escape rising frontier prices. It lands days after Mira Murati's lab shipped its first open model and while DeepSeek chases a $71 billion valuation.

Moonshot itself is reportedly raising money at around a $31.5 billion valuation, according to the Financial Times. Whether K3 lives up to its own benchmark sheet is a question the weights release will settle. For now it is the biggest open model in existence, and it is Chinese.