Listen: A free 300M model predicts Gemini's embeddings at 0.83 cosine
Mapping Gemma embeddings to Gemini embeddings via a linear transform recovers entity vector structures with high fidelity without calling the paid API.
Transcript
Imagine taking a word or phrase and feeding it to Google’s free, local Gemma embedding model. If you multiply the resulting vector by a single, simple matrix, the output is eighty-three percent similar to what you would get from the paid Gemini API. No API calls, no cost, even on data the matrix has never seen. At first glance, these two AI models look completely unrelated. But they are not. Their coordinate systems are simply rotated. Once you align them, you find they share a massive core of meaning. Specifically, they share about three hundred and forty dimensions of semantic space. They agree most on proper names and diverge on rare, technical terms. There is a catch. This map is not perfect for exact, top-ten search retrieval. The embedding space is so crowded that even a tiny angular error reshuffles the search results. But the predicted vector still lands in the exact target neighborhood nearly every single time. Ultimately, this means you can read a closed model through its open sibling. For clustering, similarity, and structural analysis, the free, local model can stand in for the paid one, showing us that different AI models are being pulled toward a common geometry of meaning.