Orders what recall finds so the right memory comes first.
- Parameters: 114M
- Context: 8,192 tokens
- Languages: 8
- Runs on: CPU or GPU, stock TEI
| English reranking | Multilingual reranking | Memory reranking | Long inputs | Serving on CPU | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Model | Overall mean | NanoBEIR | BRIGHT | MIRACL | Multilingual bank | KnowMeBench | MemBench | ConvoMem | Maximum input tokens | Rerank 50 memories ms | Rerank 50 chunks ms |
 hs-rerank-1 Hindsight · 114M | 56.0 | 71.3 | 15.1 | 57.7 | 51.5 | 52.3 | 77.1 | 66.7 | 8,192 | 1,459.8 | 21,252.8 |
 gte-reranker-modernbert-base Alibaba · 149M | 51.3 | 71.0 | 11.2 | 39.1 | 41.6 | 57.3 | 72.8 | 66.3 | 8,192 | 9,344.2 | 133,384.4 |
 bge-reranker-base BAAI · 278M | 50.4 | 61.4 | 9.5 | 51.1 | 46.8 | 50.0 | 70.5 | 63.6 | 512 | 7,870.9 | 58,415.6 |
 mxbai-rerank-xsmall-v1 Mixedbread · 71M | 48.1 | 62.9 | 11.5 | 47.9 | 42.7 | 46.6 | 64.0 | 61.4 | 512 | 2,685 | 40,870.1 |
 ms-marco-MiniLM-L12-v2 SBERT · 33M | 43.3 | 65.8 | 8.5 | 21.5 | 26.7 | 46.0 | 72.3 | 62.1 | 512 | 1,820.1 | 17,951.9 |
ms-marco-MiniLM-L6-v2 SBERT · 23M | 44.1 | 65.4 | 12.1 | 20.9 | 26.1 | 47.4 | 74.6 | 61.9 | 512 | 915.7 | 9,036.6 |
Quality vs. speed across 6 models
Top left is best value. Hover a point for details.
Model, as its provider’s logoBest value: no other model is both faster and better
Speed from the Rerank 50 memories column, on CPU.
Benchmarks
Scores are shown 0 to 100, higher is better; speed and input size as measured, with lower latency better. A shaded, outlined cell is the best in its column: blue when it is ours. Overall is the mean over the score columns.
- NanoBEIRnDCG@10, BM25 top 100, 13 tasks
- BRIGHTnDCG@10, BM25 top 100, 12 reasoning-heavy tasks
- MIRACLnDCG@10, 6 languages
- Multilingual banknDCG@10, MIRACL, MLDR, WebFAQ, mMARCO in 8 languages, 26 tasks
- KnowMeBenchnDCG@10, KaLM-Embedding top 100, 2,162 questions
- MemBenchnDCG@10, KaLM-Embedding top 100, 8,940 scenes
- ConvoMemnDCG@10, KaLM-Embedding top 100, 4,967 scenes
- Maximum inputTokens per query and document pair, from each model's configuration
- Rerank 50 memoriesMedian ms per query, 2 CPU threads
- Rerank 50 chunksMedian ms per query, 2 CPU threads