Six Ettin Rerankers Released, From 17M to 1B Parameters, With Data and Training Recipe

Tom Aarsen published six new Sentence Transformers CrossEncoder rerankers on May 19, 2026, built on Ettin ModernBERT encoders and accompanied by the full training data and recipe. The release targets retrieval pipelines by providing pointwise cross‑encoder rerankers that rescoring systems can drop into to re-score (query, document) pairs with full joint attention; this matters because it offers ready — made, reproducible options for teams balancing accuracy, latency, and cost.

The package includes six checkpoints — cross — encoder/ettin — reranker-17m-v1, cross — encoder/ettin — reranker-32m-v1, cross — encoder/ettin — reranker-68m-v1, cross — encoder/ettin — reranker-150m-v1, cross — encoder/ettin — reranker-400m-v1 and cross — encoder/ettin — reranker-1b-v1—which the author describes as state‑of‑the‑art at their respective sizes. Alongside the weights, Aarsen published the reranker dataset (cross — encoder/ettin — reranker-v1-data) and the full distillation and training configuration needed to reproduce the checkpoints.

Architecturally these are pointwise cross‑encoders that score a single (query, document) pair by letting the two texts attend to each other through all transformer layers. Because that joint encoding improves accuracy at the cost of compute, the recommended production pattern is retrieve‑then‑rerank: use an embedder to fetch top‑K candidates, then apply an Ettin reranker to reorder those K. The blog pairs the new rerankers with google/embeddinggemma‑300m on MTEB(eng, v2) Retrieval and reports benchmarked results, with additional embedder pairings provided in the release notes.

Training used a distillation recipe: a pointwise mean‑squared error (MSE) loss on scores produced by mixedbread — ai/mxbai — rerank-large-v2 evaluated over cross — encoder/ettin — reranker-v1-data. That dataset is described as a subset of lightonai/embeddings — pre-training blended with a reranked slice of lightonai/embeddings — fine-tuning. The release also includes dataset references and the evaluation scripts used for the reported metrics, enabling exact reproduction of the published evaluations.

For builders the models are standard Sentence Transformers CrossEncoder objects and can be invoked with minimal code. The post shows a three‑line example using CrossEncoder('cross — encoder/ettin — reranker-32m-v1') to compute scores and a rank helper that returns sorted indices and scores; a sample predict call produced scores like [11.393298, 2.968891], illustrating the numeric output format you’ll integrate into a retrieval stack.

Practically, the family gives teams explicit tradeoffs between latency, cost, and per‑pair accuracy: choose a 17M or 32M reranker for low‑latency re‑ranking or the 1B reranker when higher per‑pair accuracy is required. The release is also aimed at reproducibility and extension — training was bootstrapped with the train — sentence-transformers Agent Skill shipped in Sentence Transformers v5.5.0 (installation: hf skills add train — sentence-transformers [--global] [--claude])—and the author provides the overall training script so practitioners can fine‑tune or distill variants for their own corpora.