Quantizing 10M EmbeddingGemma vectors to 1 bit with MRL cuts RAM
Radar: a report that 10 million EmbeddingGemma 2 vectors, quantized to 1 bit and reduced to 256 dimensions with MRL, dropped from about 30 GB to 0.4 GB of RAM, with a declared quality loss of about 5%.
This Radar item, published on October 6, 2026, reports that the author stored 10 million EmbeddingGemma 2 vectors using 1-bit TurboQuant quantization and MRL reduction to 256 dimensions. According to the report, RAM use fell from about 30 GB to 0.4 GB. The declared quality loss is about 5%, corresponding to 94.5% of results retained when rescoring is applied.
The relevant point is the concrete trade-off between memory and quality in embedding quantization for retrieval systems. The figures come from the author and have not been verified by Rota Nacional. To check them, consult the original post through the Radar reference and note the measurement method, the evaluation set used, and the rescoring configuration. This material does not describe a Rota Nacional feature.