Any-to-any multimodal models are becoming components rather than products
A mixture-of-experts system accepting image and audio inputs, shipped under Apache 2.0 on 31 July, alongside a growing set of open VLMs. The interesting shift is positioning: these are being released as parts to build with, not as applications to use.
The first generation of multimodal releases were products with a chat box. The current ones are increasingly components — architectures and weights intended to sit inside something else, released under terms that make that legally simple.
Mixture-of-experts is the enabling choice. Routing to a subset of parameters per input is what makes a large multimodal model affordable to serve, and affordability is the difference between a capability you demonstrate and one you put in a product path that runs continuously.
Any-to-any is a design stance as much as a capability claim: accept whatever modality arrives rather than requiring the caller to route. That pushes complexity into the model and out of the application, which is the correct direction if the model is to be a component.
The licence is doing quiet work here. Apache 2.0 on a component means it can be embedded without a legal review, and in a month when other labs shipped bespoke and revenue-share terms, that is a real differentiator — arguably a larger one than any benchmark on the model card.
Thinking Machines Lab — Inkling: Our Open-Weights Model → · BentoML — Multimodal AI: A Guide to Open-Source Vision Language Models →