Promised weights are not open weights
We called a model open-sourced when it was announced as open-sourced. The weights still are not published. That is our error, and it points at a measurement the field is missing.
MiniMax H3 was announced on 3 August as an open-source release with weights promised for early August. They have not appeared. The model is live in the platform API and the consumer app.
Our correction first
We reported this release as open-sourced on the strength of the announcement. That was premature. A model is open-weight when the weights are downloadable, not when a company states an intention to publish them — and that distinction is the entire practical value of the term.
None of this implies bad faith. A timeframe was given and it has not expired by much. But the announcement is not the event, and we treated it as one.
The pattern is now standard
Announce open, ship API, publish weights later. Nearly every lab running both an open and a commercial line follows that order. Individually each instance is defensible. Collectively it means the announcement date has stopped carrying information, and only the publication date does.
Which suggests the metric nobody currently tracks: the gap. Not whether a lab calls itself open, but how many days separate the endpoint from the download — and whether that number is trending up. It is trivially measurable, entirely public, and would settle a lot of arguments that are currently conducted on vibes.
The comparison is fair to run
Because some labs ship both on the same day. DeepSeek V4 Flash 0731 arrived MIT-licensed with weights on Hugging Face, and GLM-5.2 and Kimi K3 landed on MIT or near-MIT terms. Three frontier-class releases with genuinely permissive licences is a different baseline from eighteen months ago.
Licence terms used to be where open-weight releases quietly clawed back what the announcement gave: acceptable-use riders, user thresholds, field-of-use limits. That has substantially stopped, and it stopped because being most usable is the fastest route to relevance for anyone who is not most capable.
What a permissive licence does not fix
Kimi K3 needs roughly 64 H100 or B200 GPUs across eight servers to self-host. The licence removes the legal barrier and leaves the capital barrier completely intact. Open weights is a phrase that flattens those two things, and only one of them has actually been getting better.
AIBase — Visual large models receive a major open-source announcement → · Hugging Face — Kimi K3 model overview: 2.8T parameters, MXFP4 quantization, open weights → · NIST — CAISI assessment of Z.ai's GLM-5.2 →