Open weights, on purpose

KhaM-Med builds on the MedGemma family, open-weights clinical models. The word that matters is open. Open weights mean the model itself can be downloaded, inspected, and run on servers the network controls, inside the country, under the country's law. A closed frontier model, however capable, can only be reached by sending data to someone else's servers to be answered there.

For most applications that distinction is academic. For national health infrastructure it is the whole game. A model you can host is a model whose data path you govern, whose availability you own, whose cost is your electricity rather than someone's invoice. A model you can only call is, by definition, a dependency you cannot remove.

Choosing open weights over the most capable closed model is the same kind of choice as did:web over a blockchain, or Ed25519 over something exotic. It trades a little raw capability for control, for hostability, for the ability to keep the thing running without anyone's permission. It is the boring choice, and boring is the right instinct for infrastructure.

The published performance of the open clinical models is already strong, among the best open medical models, at a fraction of frontier cost, and the gap to the closed frontier is real but narrowing. It is also the wrong thing to optimize for. The right thing is a model good enough for the task that the country can actually own. KhaM-Med starts from open weights because a model you can hold in your own hands, in your own country, under your own law, is worth more than a better one you can only borrow.