Moonshot promised Kimi K3's weights by July 27, 2026. Even if they land on time, the file is roughly 1.4 terabytes at four-bit precision, which needs about eighteen 80GB accelerators just to load before serving a single request. Open, in this case, does not mean runnable.
Read the full story: Kimi K3's Weights Are Due Today. Can You Run Them? →
Transcript
Today is Moonshot's own deadline to release the weights for Kimi K3. As of this morning there's still no repo on Hugging Face. But the release date isn't the interesting number. One point four terabytes is. K3 packs two point eight trillion parameters. Mixture of experts makes it cheap to run per token, but it does nothing to make it cheap to hold, because any token might route to any expert. So all of it stays in memory. At four bit precision that's one point four terabytes. At sixteen bit, five point six. You'd need roughly eighteen eighty gigabyte accelerators just to load it. So who is this actually for? Cloud providers, researchers auditing the model, and people building smaller distillations. Not you, on your desk. Open weights at this scale means auditability, not access.