The licence is the product
Three vision models shipped in one week at similar sizes with similar capabilities and three completely different permissions. The permissions will decide which one anyone builds on.
Everything about it says deployment except the licence
Small, multilingual, instruction-tuned. That is a specification written for production use in markets that are not English-first. Research-only means none of that use is permitted.
There is a coherent strategy behind it. Research licences establish a model in evaluations, papers and academic mindshare while keeping commercial terms open to negotiate. What they cannot do is build an ecosystem, because nobody invests engineering effort in a dependency they are forbidden to ship.
The same week, two other answers
Alibaba put a 27B with integrated vision out under Apache 2.0. LiquidAI shipped a 3B tuned for edge deployment. Cohere shipped a comparable class of model you may study and may not use.
Licence terms are now the most consequential differentiator in open-weight releases — more than parameter counts, and often more than benchmarks.
Every lab is running a different experiment about how much freedom is required to buy adoption, and the results of those experiments will be visible in eighteen months as ecosystem depth rather than as leaderboard position.
Where permissiveness collides with policy
An open release has pushed multimodal generation to 2K high-definition with synchronised audio. The capability is not novel; closed platforms have had it for months. The distribution is.
A hosted generator is a service, and services carry policy — rate limits, content filters, watermarking, terms of use, an account that can be revoked. Open weights on your own hardware carry none of that, because there is nobody in the loop to impose it.
What that does to provenance
Watermarking-by-default and the EU's synthetic-media labelling obligations both assume a generation step somebody controls. Open weights remove that step. The obligation still binds whoever publishes the output, but the enforcement point moves from the model provider to the publisher — and there are a great many more publishers than providers.
The operational conclusion for anyone building on synthetic media: do not rely on generation-side provenance surviving. Detection and disclosure are your problem, because upstream there is increasingly nobody to hand them to.
The Open Weights — Open-source AI, tracked daily → · AIBase — Visual Large Models Receive a Major Open-Source Announcement → · Let's Data Science — Multimodal AI News: Image, Video & Vision Models →