NeuML Achieves 89% of MiniLM Performance with 90K Parameter Atto Model
·2026.11.22 11:30
Key point
NeuML has released 34 ultra-small search embedding models, Atto, ranging from 840 to 28M parameters, under the Apache 2.0 license.
Details
- Released 34 ultra-small search embedding models, Atto, ranging from 840 to 28M parameters on Hugging Face under the Apache 2.0 license
- The 90K parameter model achieves approximately 89% of MiniLM's performance, while the 28M model exceeds MiniLM with an nDCG@10 of 0.5138
- Maximizes parameter efficiency by applying an architecture that separates internal processing dimensions from final embedding dimensions
- Trained on English Wikipedia, it cannot process Korean, and performance drops significantly if query/document prefixes are not applied
- The 8192-dimensional model requires 33GB of storage for indexing 1 million items, making storage costs far greater than the model file size
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.