Served on MixpeekRerankingRecommendedVerified 2026-11-05
BGE Reranker v2 M3
BAAI/bge-reranker-v2-m3
BGE Reranker v2 M3 is a text reranking model (Apache-2.0). Served on Mixpeek, scaling to zero when idle.
A Ray Serve app in Mixpeek's inference registry hosts it. It starts at zero replicas and wakes on first use.
New opens the Designer in Studio. For an existing namespace, open it in Studio and use its Designer tab.
No demo for BGE Reranker v2 M3 yet
A video, image, slideshow or playground goes here once we record one on our own data.
What it is good and bad at
- Best at
- Re-ordering a short candidate list of text documents against a query (cross-encoder, multilingual), most useful where the dense ordering is weak.
- Worst at
- Large candidate pools on CPU replicas: committed runs breached the 10 s stage ceiling, and it hurt queries whose dense ordering was already near perfect (0.986 -> 0.781 nDCG). Text only.
- Task
- Reranking
- Input
- Text
- Output
- Scores
- Use cases
- Reranking, Multilingual search
- Why we recommend it
- Served cross-encoder; raised nDCG on the TubeScience D1 eval over the unreranked baseline.
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
- pairs
- [[query, document], ...]
- text
- max_tokens
- 512
- truncation
- true
Output
- scores
- one relevance logit per pair
How Mixpeek wires it
- Extractors
- None today
- Feature URIs
- None today
- Vector indexes
- None
- Retriever stages
- rerank
- Combines with
- feature_search (E5 or any text feature) -> rerank inference_name=BAAI__bge_reranker_v2_m3 (the default)
How it runs
- Status
- Served on Mixpeek Ray Serve app in the inference registry; the rerank stage's default inference_name.
- Serve apps
BAAI__bge_reranker_v2_m3- Runtime
- transformers AutoModelForSequenceClassification + AutoTokenizer (XLMRobertaForSequenceClassification)
- Weights
- safetensors, no remote code
- Parameters
- 568 M
- Size on disk
- 2,271 MB
- Preprocessing
- Tokenize each (query, document) pair together, truncation at 512 tokens.
- Postprocessing
- Classification logit per pair as the score; sigmoid applied when the request asks for normalized scores.
- Serving
- realtime
- Hardware
- Runs on CPU (slower without a GPU)
Scale to zero
- Minimum replicas
- 0
- Maximum replicas
- 8
- Idle timeout
- 300 s
- Memory
- 2 GB
- CPUs
- 0.5
- GPUs
- 0
- Cold start, node warm
- Not recorded
- Cold start, new node
- Not recorded
Measured cost
Unmeasured
No committed throughput or cost measurement. customers/livenation/manifest-loop/README.md reports about 2 s for 10 pairs in-process and every Serve call timing out at 0.25 CPU; not a full measurement.
Our evaluations
| Dataset and metric | Score | Baseline | Date |
|---|---|---|---|
TubeScience capability parity D1, 13 scored queries nDCG (relevance by lexical overlap with the query, on the returned pool), feature_search -> rerank customers/tubescience/parity/2026-09-18-CAPABILITY-PARITY-REPORT.md | 0.9078 | feature_search dense order, same fetch: 0.8606 | 2026-09-18 |
Questions
- What is BGE Reranker v2 M3 best at?
- Re-ordering a short candidate list of text documents against a query (cross-encoder, multilingual), most useful where the dense ordering is weak.
- What are the limits of BGE Reranker v2 M3?
- Large candidate pools on CPU replicas: committed runs breached the 10 s stage ceiling, and it hurt queries whose dense ordering was already near perfect (0.986 -> 0.781 nDCG). Text only.
- Can BGE Reranker v2 M3 be used commercially?
- Yes. Apache-2.0 weights; no use restrictions.
- How does BGE Reranker v2 M3 run on Mixpeek?
- Ray Serve app in the inference registry; the rerank stage's default inference_name.