Moonshot AI published the complete weights of Kimi K3 to Hugging Face at 13:31 UTC on 27 July, eleven days after the model went live as a hosted API. The repository's commit history shows the initial upload at that minute; the five commits after it are README edits. At 2.8 trillion parameters it is the largest set of model weights ever put up for public download.

What shipped

The model card describes a mixture-of-experts architecture: 2.8 trillion total parameters with 104 billion activated per token, selected from 896 experts — 16 routed plus two shared — across 93 layers. Context length is 1,048,576 tokens. Weights ship in MXFP4 with MXFP8 activations, produced by quantisation-aware training, and the repository totals roughly 1.56 TB. A 401-million-parameter MoonViT-V2 encoder handles vision. Self-reported scores on the card include GPQA Diamond 93.5, Terminal-Bench 2.1 88.3, BrowseComp 91.2 and DeepSWE 67.5.

The licence is the story

Outlet after outlet described the terms as a "modified MIT licence." The LICENSE file in the repository is not MIT and not a modification of it. It is a bespoke Kimi K3 License — Hugging Face tags it license: other — and it carries two commercial gates. Any licensee running a model-as-a-service business whose group revenue exceeds $20 million over any consecutive 12 months must sign a separate agreement with Moonshot before using the model commercially. Any product built on it with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3" prominently in its user interface. Internal use and access through Moonshot's own products are exempt.

Open weights, not open source

Training data and training code are not included, which makes this an open-weight release rather than an open-source one — a distinction that matters more here than usual, because the terms are calibrated precisely to let students, startups and researchers use the model freely while stopping a large rival from reselling it. The permissiveness is real. It is also targeted.

The day it landed on

The timing was not quiet. Four hours before the upload, Nvidia launched an alliance arguing that open frontier models are essential to cyber defence, with no closed-model lab in it. Five hours after, Anthropic's chief executive published a statement denying his company had ever sought a ban on open weights. Washington is actively debating whether to restrict Chinese open models; the largest one now exists, in public, with a Chinese licence attached.