_by_Achim_Lammerts.jpg)
DataScience-UIBK has released an open English retrieval model that combines broad semantic matching with token-by-token matching in one search index. The team says the 179-million-parameter model scored 59.79 on BEIR, a benchmark suite for document retrieval; that result is reported by the model's authors, not independently reproduced here.
Retrieval systems rank documents against a question or search request. A dense retriever compresses each text into one vector and compares the overall meaning. A ColBERT-style system keeps vectors for individual words, letting each query word find its closest match in a document. The first can capture a broad topic; the second can reward an exact name or phrase. A common hybrid setup runs both systems and combines their scores afterward.
One index carries both matching signals
UIBK-Embedding-v1 puts two specialized heads on a mostly shared ModernBERT encoder. It places a compact representation of the whole text alongside the token vectors, then tags the two kinds so they do not match each other by mistake. ColBERT’s MaxSim score adds each query token’s best document-token match. Here, that score also includes the dense score at a weight of 0.75. The release says the model works with PyLate and a standard PLAID index, without changing that index or the scorer.
That design aims to remove a second model, a second index and a separate score-fusion step from hybrid retrieval. Late-interaction search still stores a vector for each token, so its index remains larger than a single-vector system.
Reported results and use requirements
The authors report a mean nDCG@10 of 59.79 across BEIR's 15 datasets, using the standard PyLate wrapper and default PLAID settings. nDCG@10 measures how well the first ten results are ordered by relevance. Their post says other models' scores are drawn from the official MTEB results repository; MTEB's own release notes confirm that it added this model, but do not independently validate the reported score. On a FiQA run with 57,638 documents on one H100 GPU, using 32-token queries and 300-token documents, the authors report a 580-MiB index versus 743 MiB for a two-model hybrid. Search took 7.8 milliseconds per query, compared with 4.4 milliseconds for the hybrid.
The model is English-only and was evaluated for retrieval, not other language tasks. Its authors say the weights use the Apache 2.0 license. Running it through PyLate requires the repository's remote code, plus explicit “query:” and “document:” prompts. The release is available for teams to test, but its benchmark and efficiency results remain the authors' own report.
Sources
The model release and evaluation by Abdelrahman Abdallah; UIBK-Embedding-v1 model card; MTEB 2.24.7 release notes.
Photo: Achim Lammerts (Syntaxys), Wikimedia Commons, CC BY-SA 4.0. Illustrative image only; it does not show the model or a search system.