NeuML published Atto on November 19, a family of compact retrieval models with dense, sparse and late-interaction representations. One demonstrated configuration has only 840 parameters, but it is a ColBERT-style example. It should not be described as a universal replacement for a larger model that compresses a whole document into one vector.
The authors' results cover a specified BEIR subset, rather than every task or language. An individual model card provides configuration and usage examples, making it a more precise starting point than the family name alone. We have not independently reproduced the results or evaluated a Bulgarian corpus.
Why architecture changes the budget
Comparing retrieval systems requires separating the model, stored document representations and scoring operation. Few weights do not necessarily mean a small index. If more information is retained for individual parts of a text, that information must be stored and used at query time. This is why parameter count is not a sufficient budget for the complete system.
A practical selection could start with the intended retrieval task. Is the reader looking for an exact term, a related meaning or a document containing a useful answer? For a small technical archive, a team could collect realistic questions and assess relevant documents in advance. It could then compare configurations on the same inputs instead of selecting the smallest number in a model name.
Index-building time, storage size and behavior during corpus updates should be measured separately. As a collection grows, storage may matter more than the weights. For an interactive application, query latency may be decisive. These are proposed evaluation criteria, not results from an SIExpo benchmark.
The loading example in the checked model card uses trust_remote_code. That setting is a reason to inspect the code and its version, not merely syntax to copy. We have not loaded the model or audited its implementation for this article.
Atto's interesting contribution is exploring different operating points between compactness and retrieval quality. Success in a particular application should be assessed through retrieved documents and total system cost, rather than parameter count alone. That distinction makes the development useful without turning its smallest demonstration into an exaggerated universal record.



