Is hs-rerank-1 state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.
Vectorize's 114M multilingual cross-encoder reranker, the default reranker in Hindsight 1.2.0: it scores query and document pairs of up to 8,192 tokens from 0 to 1 in eight languages. Open weights (MIT) on Hugging Face, also served on the Hindsight Inference API.
The details so far
- Lab
- Vectorize
- Announced
- Nov 3, 2026
- API
- Live · hs-rerank-1 · Hindsight Inference POST /v1/rerank (Cohere Rerank compatible), $0.015 per 1M input tokens
- Open weights
- Published · Hugging Face hindsight-ai/hs-rerank-1: FP16 safetensors and an int8 ONNX export; serves on Text Embeddings Inference on CPU or GPU
- Parameters
- 114M (113,794,177 stored parameters; 98M are the token embedding table, so each pass runs a 15M-parameter transformer)
- Type
- Multilingual cross-encoder reranker; scores each query and document pair from 0 to 1
- Base model
- jhu-clsp/mmBERT-small (fine-tuned)
- Context
- 8,192 tokens per query and document pair
- Languages
- English, Spanish, French, German, Portuguese, Japanese, Korean, and Simplified and Traditional Chinese
- License
- MIT
- Use in Hindsight
- Default reranker in Hindsight 1.2.0
- Siblings
- hs-decide-1-flash and hs-embed-1, released the same day (separate pages)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: overall mean nDCG@10 of 56.0 over seven benchmarks (NanoBEIR, BRIGHT, MIRACL, a 26-task multilingual bank, KnowMeBench, MemBench and ConvoMem) against 51.3 for gte-reranker-modernbert-base and 50.4 for bge-reranker-base.
Lab claim: median 1,459.8 ms to rerank 50 memories on two CPU threads, against 9,344.2 ms for gte-reranker-modernbert-base.
Sources
LABIntroducing Hindsight Models (Hindsight blog)hindsight.vectorize.io/blog/2026/11/03/hindsight-models · Nov 3, 2026LABVectorize announcement on Xx.com/Vectorizeio/status/2117647994048704588 · Nov 3, 2026LABhs-rerank-1 benchmarks (Hindsight Benchmarks)benchmarks.hindsight.vectorize.io/models/rerankerLABHindsight v1.2.0 release notesgithub.com/vectorize-io/hindsight/releases/tag/v1.2.0 · Nov 3, 2026REPOhindsight-ai/hs-rerank-1 on Hugging Facehuggingface.co/hindsight-ai/hs-rerank-1Tracking continues until an independent board lists it.