Hindsight AI's hs-rerank-1 Brings Multilingual Reranking to Local Machines
hs-rerank-1 is a multilingual cross-encoder reranker with 114M params, MIT licensed. Only 15M of its parameters run per forward pass; the other 98M are the embedding table. Supports 8 languages (en, es, fr, de, pt, ja, ko, zh) and 8,192-token context windows.
Takeaways
- hs-rerank-1 is a multilingual cross-encoder reranker with 114M params, MIT licensed.
- Only 15M of its parameters run per forward pass; the other 98M are the embedding table.
- Supports 8 languages (en, es, fr, de, pt, ja, ko, zh) and 8,192-token context windows.
- Built on jhu-clsp/mmBERT-small , a ModernBERT-based multilingual encoder trained on 3T tokens.
- Ships int8 ONNX for CPU and FP16 safetensors for GPU, served via Text Embeddings Inference.
- Designed as the local reranking layer for the Hindsight open-source agent memory system.
Hindsight AI has released hs-rerank-1 , an MIT-licensed multilingual cross-encoder reranker with more than 900,000 downloads on Hugging Face. The model scores a query and candidate document together, allowing a retrieval pipeline to reorder results before passing them to an application or language model. Its main engineering feature is a 114-million-parameter footprint paired with roughly 15 million parameters of transformer compute per forward pass. Hugging Face counts repeated and automated downloads, so the total does not represent unique deployments.
Most weights stay out of the matrix math
| Characteristic | Value |
|---|---|
| Total parameters | 114 million |
| Token embeddings | Approximately 98 million |
| Transformer parameters | Approximately 15 million |
| Maximum context | 8,192 tokens across the query and document |
| Documented languages | English, Spanish, French, German, Portuguese, Japanese, Korean, and Chinese |
| Deployment formats | INT8 ONNX for CPU and FP16 safetensors for GPU |
An embedding table maps token IDs to vectors by fetching the rows used in the input. The model therefore avoids multiplying through all 98 million embedding parameters during each inference call. Most dense computation happens inside the much smaller transformer, although the full parameter count still affects artifact size and memory use.
The large table comes from the jhu-clsp/mmBERT-small base model. Its 256,000-token vocabulary and 384-dimensional embeddings account for roughly 98 million parameters because 256,000 multiplied by 384 equals 98.3 million. That vocabulary provides stronger coverage for multilingual scripts, including Chinese, Japanese, and Korean, while hs-rerank-1 reduces the compute-heavy portion of the network.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
