Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

hs-rerank-1

Vectorize's 114M-parameter MIT-licensed cross-encoder, built for agent memory and fast enough to run on a CPU.

Reviewed by ReviewRamble editors · Updated

Visit hs-rerank-1 Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.8/5

A sensible replacement for the old MiniLM cross-encoders that still sit in many pipelines, small enough for a CPU, though its benchmark lead comes only from Vectorize.

What is hs-rerank-1?

hs-rerank-1 is a small multilingual reranker from Vectorize, the company behind the Hindsight agent memory server, released on 3 November 2026. It is a cross-encoder: it reads a query and a document together and returns one relevance score between 0 and 1. It is fine-tuned from mmBERT-small and covers English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese.

Of its 114 million parameters, 98 million are the token embedding table, so each pass runs a transformer of about 15 million. Inputs reach 8,192 tokens per query and document pair. It ships as FP16 safetensors and an int8 ONNX export for CPUs, and serves on Hugging Face's Text Embeddings Inference.

Since Hindsight 1.2.0 it is the default local reranker, replacing ms-marco-MiniLM-L-6-v2, which is why it passed about 2 million downloads in its first weeks. The weights are MIT-licensed; Vectorize also hosts it behind a Cohere-compatible endpoint at $0.015 per million input tokens.

How much does hs-rerank-1 cost?

PlanPriceWhat you get
WeightsFreeMIT licence
Hindsight Inference$0.015 per million input tokensCohere Rerank-compatible endpoint; output free

Free MIT weights. Hosted on Hindsight Inference at $0.015 per million input tokens (output free), or $0.02 with hs-decide-1-flash pruning; new accounts get $5 of credit.

Is hs-rerank-1 worth it?

The model's real competition is the ms-marco MiniLM cross-encoders that many pipelines still use as a default, Hindsight's own included until version 1.2.0, and against those it is a clear step up. They read 512 tokens and were trained on English MS MARCO data; hs-rerank-1 reads 8,192 tokens in eight languages and stays small enough to run inside an API process. On Vectorize's leaderboard it averages 56.0 across seven benchmarks, against 44.1 for ms-marco-MiniLM-L6-v2, 50.4 for bge-reranker-base and 51.3 for Alibaba's gte-reranker-modernbert-base.

The gains are largest outside English. On MIRACL, a multilingual set, it scores 57.7 against 20.9 for MiniLM-L6 and 39.1 for the gte model. On BRIGHT, a reasoning-heavy benchmark, every small model struggles: 15.1 is the best of the group and still low in absolute terms. Vectorize ran every comparison itself with the same settings for each model and published the method, but no outside lab has repeated it.

Speed is the reason to choose it over larger open models. On two CPU threads it reranks 50 memories in about 1.5 seconds, about a sixth of the time gte-reranker-modernbert-base takes, while MiniLM-L6 is faster still at 0.9 seconds. For 50 longer chunks the figure rises to about 21 seconds, so long documents on a CPU remain slow; a GPU or the hosted API is the answer there.

It does not compete with Cohere Rerank 5, Voyage rerank-3 or the 4B and 8B Qwen3 rerankers on hard queries, and it is not meant to. The hosted version is priced at $0.015 per million tokens, below Cohere's Fast tier at $0.02, and the hosted docs truncate documents at 4,096 tokens, half the model's own limit.

Where it earns its place

  • A developer running Hindsight as memory for a coding agent

    Rank recalled memories well without a GPU or another paid API

    Upgrade to Hindsight 1.2.0 or later, where hs-rerank-1 is already the default local reranker and downloads on first start. It runs in the API process on CPU, so the only cost is CPU time on the machine that already runs Hindsight.

  • An engineer at a European SaaS company with a MiniLM cross-encoder in its support search

    Rerank German, French and Spanish tickets as well as English ones

    Swap the model ID in Sentence Transformers' CrossEncoder or Text Embeddings Inference for hindsight-ai/hs-rerank-1. On Vectorize's MIRACL results that moves the score from 20.9 to 57.7, with latency on CPU about 1.6 times MiniLM-L6's.

  • A small team already calling Cohere Rerank from LangChain

    Cut the reranking bill for short, high-volume memory lookups

    Point the Cohere client at inference.hindsight.vectorize.io and pick hs-rerank-1 at $0.015 per million input tokens; a billion tokens a month costs $15. Expect a drop in quality on hard queries.

Before you commit

Setup
pip install sentence-transformers and load CrossEncoder("hindsight-ai/hs-rerank-1"), or start the Text Embeddings Inference CPU or GPU image with the model ID. Scores come back between 0 and 1.
Integrations
Sentence Transformers, Text Embeddings Inference, ONNX Runtime and Hindsight's local, ONNX, TEI and hosted reranker providers; the hosted endpoint works with the Cohere SDK.
Data and privacy
Self-hosted, documents never leave your machines. The hosted Hindsight Inference docs do not state retention; Vectorize publishes a Trust Center and data processing agreement.
Cost at scale
Self-hosted cost is CPU time. Hosted: $0.015 per million input tokens with output free, so 20 passages of 300 tokens per call, 2 million calls a month, is about $180.
Limits
8,192 tokens per pair self-hosted, 4,096 per document and 1,000 documents per request on the hosted API. The hosted rate limits are not published. Eight languages only.

What’s good

  • 114M parameters; runs on a CPU with an int8 ONNX export
  • 8,192 tokens per pair in eight languages
  • MIT licence with no usage restrictions
  • Cohere-compatible hosted endpoint at $0.015 per million tokens

What’s not

  • All benchmarks are Vectorize's own
  • Weak on reasoning-heavy queries (15.1 on BRIGHT)
  • Slow on long chunks without a GPU
  • Hosted API truncates documents at 4,096 tokens

Best for

Developers who want a small, permissively licensed multilingual reranker for agent memory or CPU-only RAG in place of an old MiniLM cross-encoder.

hs-rerank-1 alternatives

hs-rerank-1: common questions

Is hs-rerank-1 worth it?
A sensible replacement for the old MiniLM cross-encoders that still sit in many pipelines, small enough for a CPU, though its benchmark lead comes only from Vectorize. We rate it 3.8 out of 5.
How much does hs-rerank-1 cost?
Free MIT weights. Hosted on Hindsight Inference at $0.015 per million input tokens (output free), or $0.02 with hs-decide-1-flash pruning; new accounts get $5 of credit.
What are the downsides of hs-rerank-1?
All benchmarks are Vectorize's own; weak on reasoning-heavy queries (15.1 on BRIGHT); slow on long chunks without a GPU; hosted API truncates documents at 4,096 tokens.
Who is hs-rerank-1 best for?
Developers who want a small, permissively licensed multilingual reranker for agent memory or CPU-only RAG in place of an old MiniLM cross-encoder.
How much does hs-rerank-1 cost for a team or in production?
Self-hosted cost is CPU time. Hosted: $0.015 per million input tokens with output free, so 20 passages of 300 tokens per call, 2 million calls a month, is about $180.
Is hs-rerank-1 safe for company data?
Self-hosted, documents never leave your machines. The hosted Hindsight Inference docs do not state retention; Vectorize publishes a Trust Center and data processing agreement.
What are the best hs-rerank-1 alternatives?
The closest alternatives we have reviewed are mxbai-rerank v2, Qwen3-Embedding and Qwen3-Reranker, Jina Reranker v3.5 and Cohere Rerank 5.

Answers checked Nov 20, 2026.