MIT-licensed frontier weights stop being remarkable
DeepSeek V4 Flash 0731 ships MIT. GLM-5.2 shipped MIT with no regional limits. Kimi K3 shipped under a modified MIT permissive enough for commercial use. Three of the most capable open-weight models available now carry among the most permissive licences in software — a position that would have been startling eighteen months ago.
Licence terms used to be where open-weight releases quietly clawed back what the announcement gave away: acceptable-use riders, user-count thresholds, field-of-use restrictions. Three consecutive frontier-class releases landing on MIT or near-MIT changes the default expectation, and defaults are what most downstream builders actually rely on.
The competitive logic is not mysterious. If your model is not the most capable available, the fastest route to relevance is being the most usable — and licence friction is the cheapest friction to remove. It costs nothing to grant and it is immediately legible to the developers deciding what to build on.
Worth naming what remains hard regardless of licence. Kimi K3 needs roughly 64 H100 or B200 GPUs across eight servers to self-host. A permissive licence removes the legal barrier and leaves the capital one entirely intact, which is a distinction the phrase open weights tends to flatten.
NIST — CAISI assessment of Z.ai's GLM-5.2 → · Hugging Face — Kimi K3 model overview: 2.8T parameters, MXFP4 quantization, open weights → · DataNorth — Zhipu AI releases GLM-5.2 open-weight AI model →