Kimi K3 Quietly Arrives

Kimi does not appear to have published a launch post for K3 yet, but the model is already available through Vercel AI Gateway and documented on the Kimi API Platform.

The headline number is 2.8 trillion parameters. If Moonshot eventually releases the weights, K3 would be among the largest open-weight models. For now, that remains speculation.

Vercel lists K3 at $3 per million input tokens and $15 per million output tokens. That is frontier-tier pricing, which suggests Moonshot has high confidence in the model. We still need benchmarks—and real use—to know whether the capability justifies it.

The most interesting part to me is dynamic tool loading. K3 lets an application inject complete tool definitions into the message history only when they are needed. Appending the declaration preserves the cached prefix, so tools can be loaded progressively without invalidating the cache that came before them.

That is exactly the design constraint I wrote about yesterday in Your Agent Has Too Many Tools: tool discovery is useful only if loading a tool does not make the rest of the session unnecessarily expensive.

Kimi’s implementation means dynamic tool loading is no longer limited to Anthropic and OpenAI’s frontier APIs. I hope this becomes a standard capability across more models and providers. 🤞