The licence is the product
Apache 2.0 won. That sentence sounds like housekeeping and is actually a competitive outcome, because legal review is where an evaluation quietly dies.
Where evaluations actually die
Not in benchmarking. In legal review. Apache 2.0 and MIT are terms corporate counsel already has a cached answer for; a bespoke community licence, however reasonable, requires a fresh read by someone with a queue.
The standard-licence models get tried. The bespoke-licence model gets scheduled.
Llama's distribution is large enough to absorb that friction. But the labs that removed it removed it deliberately, and friction compounds in a market where the evaluation window is now three weeks wide.
Open weights, closed corpus
The licence race being settled leaves the harder question untouched. Apache 2.0 on a model whose training data is undisclosed grants freedom to use and no ability to audit. Those were never the same axis, and conflating them has flattered the field for years.
The last holdout is being tested
Staged release before a public drop is a normal sequence, not a betrayal. But the frontier of what that lab can do is currently visible only to people who signed something, and the eventual licence is the tell. Apache 2.0 means the early-access period was a runway. Bespoke terms mean the last major permissive holdout has joined everyone else.
What this asks of you
Treat the licence as a first-class selection criterion, ahead of a few benchmark points. Time-to-first-deployment is dominated by review, not by capability, and one major vendor is still asking you to spend that time.
And keep the two questions separate when you write your own policy. "Can we use it" is settled. "Can we inspect it" is not, and nobody's licence answers it.
Presenc AI — Open-Weight License Landscape 2026 → · TechTimes — Mistral AI Targets Frontier Gap With Open-Weight Model Entering July Early Access →