EmbeddingGemma 2

Simon Willison published a comment on EmbeddingGemma 2, referencing Hacker News, in which he said he appreciates that the model is released under the Apache 2.0 license.

Willison argued that closed, proprietary, hosted-only models are a poor fit for embedding work specifically. He wrote that most applications of embedding models involve calculating thousands or even millions of embedding vectors and storing them for later comparison. If such a model is proprietary, he said, the vendor will likely someday decide to stop offering it. A better replacement model may exist, but existing stored vectors would still need to be recalculated at a cost.

As an example, Willison noted that in April 2024 OpenAI offered to cover the financial cost of users re-embedding content with new models, and linked to OpenAI's GPT-4 API general availability post. He added that he does not think that is something that can be relied on from every provider.

Willison also wrote that he does not want to host the model himself. He said he would rather pay a provider for a hosted model while knowing that if the provider ever stops hosting it, he can run the open weights version himself or find another vendor to do so.

The comment was tagged with google, ai, generative-ai, embeddings and gemma.