moonshotai/Kimi-K3

Moonshot AI followed through on its promise and dropped the weights for Kimi K3, the successor to July's K2, and it's a monster: 2.8 trillion parameters, 1.56 terabytes sitting on Hugging Face waiting for anyone with the hardware to run it. That's the easy part of the story. The harder part is what you're actually agreeing to if you want to use it.

Back with K2, Moonshot wrote its own twist on the MIT license, adding a clause that anyone shipping a commercial product with more than 100 million monthly active users, or pulling in over $20 million a month, had to slap "Kimi K2" somewhere visible on their interface. Annoying, but survivable. K3 tightens the screws further. Moonshot no longer even calls it a modified MIT license — it's just its own thing now. And buried in the terms is a requirement that any Model-as-a-Service business generating more than $20 million in revenue over any rolling 12-month period has to go negotiate a separate deal with Moonshot before using K3 commercially at all.

Give Moonshot some credit for honesty here: nowhere in its materials does it call this "open source." The company consistently uses "open weight" instead, which is the accurate term and a small act of restraint in an industry that loves stretching the open-source label past its breaking point.

Practically speaking, none of this seems to be scaring off the API resellers. OpenRouter already lists K3 through seven different providers, and most of them are matching Moonshot's own pricing of $3 per million input tokens and $15 per million output tokens. So if you just want to call the model through an API, the licensing fine print barely touches you — it's the companies trying to build large-scale services directly on top of the weights who'll need to read the contract carefully.