Why Embedding Models Should Be Open

Simon Willison welcomes the release of EmbeddingGemma 2 under the Apache 2.0 license, and uses the occasion to state a principle: for embedding models in particular, a closed, proprietary, hosted-only option makes no sense.

The reasoning is simple and rooted in how these models are actually used. The typical scenario is to compute thousands — even millions — of embedding vectors and store them for later comparison. This is not a one-off request but a database that lives for a long time.

The proprietary trap

If the model is closed, the vendor will eventually decide to stop offering it. A better version will take its place — but all your millions of previously computed vectors will be incompatible with it. You'll have to recalculate everything, and pay for it again.

Willison recalls the one encouraging precedent: in April 2024 OpenAI offered to cover the financial cost of users re-embedding their content when moving to new models. But, he notes, relying on such generosity from every provider would be naive.

Open weights as a fallback

Tellingly, the author is not urging everyone to self-host. On the contrary, he is happy to pay for convenient hosting. The difference is that with open weights he keeps a fallback plan: if a provider ever shuts the service down, he can run the model himself — or find another vendor to do it for him.

In effect, the open license works here not as an ideological statement but as an insurance policy. You enjoy the convenience of the cloud for exactly as long as it exists, but your stored data never becomes hostage to someone else's business decision.

Source: Simon Willison