gte-reranker-modernbert-base

gte-reranker-modernbert-base, published by Alibaba's Tongyi Lab on 20 January 2025, is the second half of a two-model pipeline and the more heavily used one. The first half, an embedding model, compresses every document into a vector once and retrieves a rough shortlist fast. This model then reads the query and each candidate document together, in full, and scores the match — accurate but far too slow to run against a whole index, which is why it only ever sees the top twenty or hundred results. It is 149 million parameters over 22 layers and takes 8,192 tokens, the same encoder-only shape as its sibling. Both are fine-tunes of ModernBERT-base, released in December 2024 by Answer.AI and LightOn; Alibaba contributed the retrieval training, not the backbone. Unlike the embedding model it emits no vector at all, only a relevance score, so it cannot be used to build an index. The adoption gap between the two is the striking part. In the 30 days to 23 September 2026 this model was downloaded 2,544,278 times against 158,273 for the embedding model published the same day — sixteen to one — while collecting fewer than half the likes (98 against 201). The pattern fits how retrieval systems are actually assembled: teams keep whichever embedding model their index was built with, because changing it means re-indexing everything, but a reranker can be swapped in at the end of an existing pipeline in an afternoon. On vendor figures it scores 56.19 on BEIR, 90.68 on the long-context LoCo suite and 79.99 on the code-retrieval CoIR suite, improving on the embedding model in every case — LoCo by three points. One caveat readers should know: the summary table and the detailed table in the same model card disagree about BEIR. Adding up the fifteen individual BEIR scores the vendor lists gives an average of 56.74, not the 56.19 printed in the summary. The same arithmetic on the embedding model's row reproduces its published 55.33 exactly, so the discrepancy is confined to this model, and the headline figure is the more conservative of the two. Weights are Apache 2.0, with no added restrictions.

▸Videos

No videos yet.

▸Reviews