Kimi K3 was announced on July 16th as the largest open-weight model ever released. I used to read a release like that through a simple split—open or closed, protect the weights or give them away. It isn't that simple anymore.
Anthropic, for one, seems to be running three plays at once: closed on the flagship model, open on the protocol (MCP, now the industry standard for building agents, adopted by AWS, Microsoft, everyone), and building a product layer of industry-defining patterns around it—Claude Code, Claude Cowork, and the primitives they run on (skills, harnesses)—a whole way of working that reaches well past the model itself. The protocol can be given away because it was never the thing being sold, and the further it travels, the more valuable the model behind it becomes. Google is behind A2A, OpenAI and Stripe behind ACP: everyone deciding, layer by layer, which part to open and which to hold.
Underneath all that deciding there's a movement nobody quite controls, and the word for it is "distillation."
Right now it's mostly thrown as an accusation—Anthropic charged Moonshot in February with siphoning Claude's reasoning traces to train Kimi on the cheap—and that may be true in part, but it mishears what the word means. Distillation isn't just copying. In its older sense it's concentration—you drive off the volume and keep what matters, and what's left can be more potent than what you started with. The alchemists called that the quintessence, the fifth essence you get once the four ordinary ones have boiled away.
And distillation in that sense is the whole story of this field. In 2017 a single paper distilled years of scattered research intuition into one architecture and called it "attention." The models distilled the open web, the books, the code into a few hundred gigabytes of weights. OpenAI distilled the model into a product, and that product got distilled in turn by a Cambrian explosion of competitors who shipped the same shape within a year. The workflows locked inside SaaS applications, the captive logic that was supposed to be their moat, are being distilled out into agentic workbenches that leave the old interfaces behind.
Each turn moves the essence up a layer, and the moat moves with it—which is what all the open-versus-closed maneuvering is really chasing. But you only get to decide the part you open. Kimi is the reminder that the closed lab's own advantage, the thing it guarded most carefully, is just the feedstock for the next distillation—cheaper, open, and not necessarily worse than what it was distilled from. You decide what you open; you don't decide what gets distilled out of you.
