Data as well as weights
Releasing a training corpus is rarer than releasing a model, and for reasons that are mostly legal rather than technical. It is also the only version that supports the scientific claim.
What each one buys you
Weights let you run and fine-tune. Data lets you understand why the model behaves as it does, reproduce the training, and audit what went in. Only the second supports the reproducibility argument that open release is usually justified by.
Which is why almost nobody does it. Corpora carry licensing exposure, provenance questions and competitive information, and most labs have decided the risk outweighs the benefit.
Grounding is the capability that matters here
Point-driven reference across single images, multiple images and video is what connects language to specific locations in a scene. It is the difference between a model that describes an image and one that can be instructed about it.
Which is exactly the capability underneath robots taking spoken work orders. "The pallets on aisle four" is a grounding problem before it is a robotics problem.
Weights, data, licence. Three axes, one adjective.
The component turn
Multimodal models are increasingly released as parts to build with rather than applications to use — mixture-of-experts for serving cost, any-to-any input so the caller does not route, permissive licences so nothing needs a legal review.
In that context the licence may be a bigger differentiator than any benchmark on the model card. A component nobody can embed is not a component.
arXiv — Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding → · BentoML — Multimodal AI: The Best Open-Source Vision Language Models in 2026 → · Thinking Machines Lab — Inkling: Our Open-Weights Model →