Six new open-source AI rerankers just dropped, hitting state-of-the-art accuracy at every size

A new family of six open-source cross-encoder rerankers, all built on the ModernBERT architecture, is now available and claims state-of-the-art performance for their respective model sizes. The models range from a tiny 17 million parameters up to a robust 1 billion parameters, all trained using the same distillation recipe and released with full training data and code.

This is a big deal for anyone building retrieval-augmented generation (RAG) systems because a reranker is the secret sauce that makes search results actually good. As the developer behind the project points out, an embedding model retrieves candidates quickly by comparing separate vectors, but a cross-encoder like these reads the query and document together to produce a deeply accurate relevance score. The trade-off is speed, which is why the standard “retrieve-then-rerank” pattern exists: use a cheap embedder to pull the top 100 candidates, then let the cross-encoder re-order them.

These Ettin rerankers were distilled from the powerful mxbai-rerank-large-v2 model on a massive, curated dataset. The training recipe was bootstrapped using a new AI agent skill in Sentence Transformers v5.5.0, hinting at a future where fine-tuning these models becomes a casual conversation with your coding agent. The models support up to 8,000 tokens of context, which ModernBERT’s long-context pre-training makes practical, and can be plugged in with just three lines of Python. If you want pure throughput, activating Flash Attention 2 and bfloat16 can yield speedups between 1.7x and 8.3x depending on the model size.

The real test is how they perform in a two-stage system. When paired with a mid-range embedder like google/embeddinggemma-300m, the tiny 17M Ettin reranker can boost retrieval performance substantially. The point isn’t just raw benchmarks—it’s that you no longer have to choose between a model that’s fast and one that’s accurate. You can pair a lightweight embedder with a small cross-encoder and get results that previously required a massive, expensive model. For teams watching their infrastructure costs, that’s not just academic.