SAVRN Model Hub · Comparisons
bge-reranker-base vs roberta-large-mnli
Bge-reranker-base has 278M parameters and roberta-large-mnli has 356M parameters; both are released under MIT License; at 16-bit, bge-reranker-base needs about 0.7 GB (1x MI300X from $1.85 an hour) and roberta-large-mnli about 0.9 GB (1x MI300X from $1.85 an hour).
| Field | bge-reranker-base BAAI/bge-reranker-base | roberta-large-mnli FacebookAI/roberta-large-mnli |
|---|---|---|
| Publisher | Beijing Academy of Artificial Intelligence | Facebook AI community |
| Task | Text classification | Text classification |
| Modality | Text | Text |
| Parameters, as reported | 278M parameters | 356M parameters |
| Architecture | XLMRobertaForSequenceClassification | RobertaForSequenceClassification |
| Library | sentence-transformers | transformers |
| Context length | 514 tokens | 514 tokens |
| Repository size | 3.4 GB | 5.7 GB |
| Artifact formats | safetensors, onnx, pytorch | safetensors, pytorch, jax, tf |
| License | mit | mit |
| Access | Open weights, no gate | Open weights, no gate |
| Memory at 16-bit (weights and margin) | 0.7 GB | 0.9 GB |
| Cheapest GPUs at 16-bit, per hour | 1x MI300X, $1.85 | 1x MI300X, $1.85 |
| Memory at 4-bit (weights and margin) | 0.2 GB | 0.2 GB |
| Cheapest GPUs at 4-bit, per hour | 1x MI300X, $1.85 | 1x MI300X, $1.85 |
| Revision viewed | 2cfc18c9415c | 2a8f12d27941 |
| Downloads reported by the hub | 4M | 238.7k |
| Last observed | 2026-09-19 | 2026-09-18 |
An evaluation row appears only where at least two of these models report the same benchmark with the same stated configuration, metric, unit and setup. Different evaluators stay named in each cell. Values are shown as reported: no unit conversion, no ranking.
Other Reported Results
These results are listed for each model on its own, because the conditions needed to compare them are not stated or do not match. Two results that leave a condition blank are not assumed to share it.
bge-reranker-base
| Benchmark | Conditions | Result | Reported by | Revision | Date |
|---|---|---|---|---|---|
| MTEB CMedQAv1 | Configuration defaultTask RerankingMetric mapComparison conditions not established | 81.2721 | BAAI Publisher reported |
Evaluated revision not stated | — |
| MTEB CMedQAv1 | Configuration defaultTask RerankingMetric mrrComparison conditions not established | 84.1424 | BAAI Publisher reported |
Evaluated revision not stated | — |
| MTEB CMedQAv2 | Configuration defaultTask RerankingMetric mapComparison conditions not established | 84.1037 | BAAI Publisher reported |
Evaluated revision not stated | — |
| MTEB CMedQAv2 | Configuration defaultTask RerankingMetric mrrComparison conditions not established | 86.7938 | BAAI Publisher reported |
Evaluated revision not stated | — |
| MTEB MMarcoReranking | Configuration defaultTask RerankingMetric mapComparison conditions not established | 35.4601 | BAAI Publisher reported |
Evaluated revision not stated | — |
| MTEB MMarcoReranking | Configuration defaultTask RerankingMetric mrrComparison conditions not established | 34.6024 | BAAI Publisher reported |
Evaluated revision not stated | — |
| MTEB T2Reranking | Configuration defaultTask RerankingMetric mapComparison conditions not established | 67.2773 | BAAI Publisher reported |
Evaluated revision not stated | — |
| MTEB T2Reranking | Configuration defaultTask RerankingMetric mrrComparison conditions not established | 77.1315 | BAAI Publisher reported |
Evaluated revision not stated | — |
SAVRN's Notes on bge-reranker-base
Retrieval pipelines need a second pass that re-scores what the first-stage retriever returns, and that is this 278M-parameter classifier's job: it reads a query and a passage together and scores the match. Memory is not the decision. At 16-bit the weights are 0.6 GB and the run needs 0.7 GB, under one percent of the 192 GB MI300X we price at $1.85 an hour, so it never gets its own card; it rides beside whatever generation model lives there.
The MIT terms are short and permissive: commercial use, modification and redistribution are allowed provided the copyright and permission notices stay with the files. The 514-token context must hold query and passage together, so chunk to fit. And the publisher's summary says newer rerankers with larger inputs and more languages shipped March 18, 2024, so decide whether this version or a successor is what you standardize on.
SAVRN's Notes on roberta-large-mnli
This is RoBERTa large fine-tuned on MultiNLI: it decides whether one sentence entails, contradicts or is neutral toward another. For years it was the standard engine behind zero-shot classification pipelines, and it has 356M parameters.
The MIT license allows commercial use. Newer NLI models in our catalog, especially the DeBERTa-v3 ones trained on several NLI datasets, are generally more accurate. Keep this one where an existing system depends on its behavior.
Questions
Which is larger, bge-reranker-base or roberta-large-mnli?
roberta-large-mnli (356M parameters) is larger than bge-reranker-base (278M parameters), by the parameter counts their publishers report.
Which is cheaper to run, bge-reranker-base or roberta-large-mnli?
At 4-bit, bge-reranker-base fits on 1x MI300X from $1.85 an hour and roberta-large-mnli on 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use bge-reranker-base commercially?
Yes. bge-reranker-base is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.
Can I use roberta-large-mnli commercially?
Yes. roberta-large-mnli is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.