The biggest open model ever lands today! But…
The biggest open model to date lands today but you still can't run it. Moonshot AI publishes Kimi K3's full weights today, 27th Jul’26; a 2.8-trillion-parameter model, the largest open-weight release anyone has shipped so far. Even squeezed to 4-bit precision the weights need roughly 1.4 terabytes of fast memory just to sit resident, before the model reads a single token of your prompt. That points at racks of the newest Nvidia Blackwell or AMD MI400 silicon; a full top-end eight-card server node (about 1.5TB) barely fits the weights. So whether you can legally use it in a product is unconfirmed until the text lands with the weights.
What it actually means
'Open' and 'runnable' have quietly come apart at the top of the market. Downloadable weights sound like control; no per-token bill set by one vendor, no capability deprecated out from under a shipping product. However, at 1.4 terabytes, the only operators who can actually host K3 are the clouds and inference providers already running that hardware. Most teams who 'adopt' it will rent it through someone else's infrastructure and inherit a version of the same provider dependency they were trying to leave. For the 'own it vs rent it' question this Digest has circled all month, this is the honest answer: at the frontier, owning is real on paper and rented in practice.
'Open weights' at 1.4 terabytes isn't a model you run. It's a model you rent from whoever can afford to keep it warm.
Source: TECHi · Monday 27 July 2026