Moonshot to open-source Kimi K3 weights
Moonshot says it will publish Kimi K3's full 2.8T-parameter weights July 27, eleven days after an API-only launch, the largest open-weight release yet.
Moonshot AI’s Hugging Face page for Kimi K3 is now live and marked “upcoming release,” backing up a promise the company made when K3 launched API-only on July 16: full weights would follow by July 27. If it ships on schedule, K3 becomes the largest open-weight model released to date, a sharp reversal from a launch that broke with every prior Kimi model by shipping closed.
The numbers are steep. K3 is a roughly 2.8-trillion-parameter mixture-of-experts model, and reporting on the release puts the download at about 1.4 terabytes in MXFP4, a 4-bit floating-point format with per-block scaling, versus roughly 5.6 terabytes at full 16-bit precision. Loading it takes on the order of eighteen 80GB-class accelerators, or a single 8-card node of newer 192GB parts, the kind of hardware only cloud providers and inference platforms run racks of. There is no realistic path to running this on a workstation.
One detail is still missing: Moonshot has not published K3’s license, saying the terms will land with the weights themselves. Independent analyst Nathan Lambert, who has tracked the release, frames his own analysis as conditional on “the assumption that Moonshot keeps their promise,” a fair caveat given K3 already shipped later than teased once before.
What it means for operators
Don’t route production workloads to a self-hosted K3 on the strength of pre-release coverage. Wait for Moonshot’s own model card and license text, confirm your inference stack actually supports MXFP4 at this scale, and treat every benchmark claim as provisional until independent evals land, the same caution that applied to K3’s API launch and to Moonshot’s own signup freeze days later. If the release holds, it reopens the question of whether frontier-scale open weights remain a Chinese-lab specialty that US labs aren’t matching at this parameter count.