Why does the license matter for embedding models?

Embedding applications typically calculate thousands or millions of vectors and store them. If a closed model is retired, users must pay to recalculate all stored vectors with a replacement.

Does Willison want to host the model himself?

No. He says he would rather pay a provider for a hosted model, while knowing he could run the open weights version himself or switch vendors if hosting stops.

What did OpenAI offer in April 2024?

OpenAI offered to cover the financial cost of users re-embedding content with its new models, but Willison says that is not something to rely on from every provider.

Large Language ModelsOpen-Source AIAI Business & IndustrySimon Willison's Weblog6h agoOct 7, 2026, 06:00 JST

EmbeddingGemma 2 ships under Apache 2.0, Willison says

Send on LINE

3 Key Points

What happened
Simon Willison praised EmbeddingGemma 2 for carrying the Apache 2.0 license, noting embedding models are poorly suited to closed, hosted-only offerings, since switching models means paying to recalculate millions of stored vectors.
Why it matters
Embedding applications typically store thousands to millions of vectors, so an open license lets users keep the option of running the model themselves or finding another host if a vendor discontinues it.
What to watch
Willison still prefers paying a provider to host the model, so the value hinges on open weights remaining available. He noted OpenAI's April 2024 offer to cover re-embedding costs is not something to rely on from every provider.

WHO IT HITSThis lands on developers and product teams whose applications store large numbers of embedding vectors, and who would otherwise face re-embedding costs if a hosted model is retired.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The post contrasts EmbeddingGemma 2's Apache 2.0 license with the risks of closed, hosted-only embedding models. Willison notes that embedding workflows tend to involve calculating thousands or even millions of vectors that are stored for later comparison. If the model behind those vectors is proprietary, a vendor may eventually retire it in favor of a better model, leaving users to pay to recalculate the existing stored vectors.

He cites OpenAI's April 2024 offer to cover the financial cost of users re-embedding content with its new models, but writes that this is not something to rely on from every provider. His stated preference is to pay a provider for a hosted model while retaining the fallback of running the open weights version himself or finding another vendor. The value of the Apache 2.0 release therefore hinges less on hosting convenience than on keeping that fallback available.

FAQ

Why does the license matter for embedding models?: Embedding applications typically calculate thousands or millions of vectors and store them. If a closed model is retired, users must pay to recalculate all stored vectors with a replacement.
Does Willison want to host the model himself?: No. He says he would rather pay a provider for a hosted model, while knowing he could run the open weights version himself or switch vendors if hosting stops.
What did OpenAI offer in April 2024?: OpenAI offered to cover the financial cost of users re-embedding content with its new models, but Willison says that is not something to rely on from every provider.

Simon Willison's WeblogRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.