Dzen open-sources dzen_embedder: local embeddings for RAG 64x cheaper than OpenAI API

💻 Dzen open-sources embedder for RAG — 64x cheaper than OpenAI

Dzen (dzen.dev) has open-sourced dzen_embedder: a service that downloads a chosen model from Hugging Face and serves an OpenAI-compatible /v1/embeddings endpoint. On a GTX 1080 — about 248 500-token chunks per second.

🌍 Embedding costs are a rare AI expense category where self-hosted hardware beats APIs on price: a server ≈ $100/month vs. $6,400/month at OpenAI (text-embedding-small) and $3,900/month at Cloudflare Workers AI (bge-m3) for 640 million chunks per month. For high-volume RAG, self-hosting with a compatible API becomes the default.

👤 Installs in an evening on any GPU or Mac with Apple Silicon: you change base_url in your code, but you should check the quality of the local model on your own data. Data doesn't leave your infrastructure, tests with up to 64 parallel connections passed without errors. License: 0BSD.

Source 1: https://dzen.dev/blog/local-dzen-embedding-for-rag/ Source 2: https://github.com/dzenplatform/embedder