EmbeddingGemma 2

EmbeddingGemma 2 comes with an Apache 2.0 license, and that is the part that stands out here. For embedding models, open weights are not some abstract ideology; they are insurance against getting trapped by a provider’s future product plans.

That matters because these models are often used at huge scale. A real deployment can mean thousands or even millions of embedding vectors, all stored for later comparison. If the model is closed and hosted-only, then a provider can eventually stop offering it, swap in something newer, and leave customers with a bill for recalculating everything they already stored.

The source points to OpenAI’s April 2024 promise to cover the financial cost of re-embedding with new models, but treats that as an exception rather than a rule. The broader point is simpler: don’t build on the hope that every vendor will be generous when the model changes.

There is a nice middle ground here. A team can still pay someone to host the model, and still avoid lock-in if the weights are open. If that host disappears, the open version can be run elsewhere, which is exactly the kind of boring freedom infrastructure should have.