Multimodal AI leans native, but shipping products are still mostly assembled — resolves 2027-07-14

AI that sees images, hears audio and makes pictures can be built two ways: by bolting an image-reading model, an image generator and voice tools together, or by training one model from the start to take in and put out every format. The route decides whose products behave consistently and feel smooth, and whether the open-source community can keep up. People who train frontier models mostly see the native route as technically better, but mainstream assistants today are assembled.

SecondSource’s reading: the contest is still open and leans toward the native route (early fusion, putting every format into one model during training), but it is not settled.

Shipping lags, especially in open source: mainstream assistants are still mostly an image-reading language model plus bolt-on image and voice tools, and open-source efforts to copy the native route are clearly behind.

One public example of the native route is Chameleon, a family of early-fusion models Meta FAIR released in May 2024.

Chameleon Team — Meta FAIR — arXiv
“We present Chameleon, a family of early-fusion…”

Reviewed July 2027: if mainstream assistants are still mostly a VLM plus bolt-on components, the lean has not paid off. Two or more frontier flagships going natively any-to-any confirms it.

What would overturn it

Open source reproducing native any-to-any at matching quality settles the question. The assembled route continuing to win decisively on cost reverses the lean.