EmbeddingGemma 2

Blog · Watch/Read · (gathered 2026-10-14)

• Simon Willison advocates for open-licensed embedding models (e.g., Apache 2.0 for EmbeddingGemma 2) over proprietary, hosted-only alternatives, emphasizing their necessity for long-term application stability.
• He argues that proprietary models create significant vendor lock-in and risk, as providers can discontinue models, forcing users to re-calculate potentially millions of existing embedding vectors at considerable cost, a risk not reliably mitigated by vendor offers (e.g., OpenAI’s past re-embedding cost coverage).
• The practical takeaway is to choose hosted embedding services that utilize open-weights models, providing the crucial fallback option to self-host or switch to another vendor without needing to re-embed all data if the original service ceases.