dev split. Real numbers from the tuning set, not the held-out test split; retrieval only, no answer layer has run yet.

Model selection

Every embedder and reranker on the shortlist is scored on a public Czech retrieval pack before the DealBench dev split decides. Only licences that permit commercial use can become a pick; others are measured and shown as research-only. Scores are the slice's MTEB main metric, in points.

Pick rule. Highest mean main score over the stage's slices (all slices for embedders, reranking slices for rerankers), among models with a licence that allows commercial use and a full (not subsampled) score on every one of them; ties go to the smaller model. No pick is made with fewer than two such contenders. Provisional until confirmed on the DealBench dev split. Generated 2026-10-01T00:57:24+00:00.

Dense embedders

decided by the mean over these slices
Dense embedders candidates
ModelLicenceHostingVRAMbelebele-cs-q-en-doc nDCG@10wiki-cs-rerank MAPMeanStatus
bge-m3 Baseline (D1); no query instruction.MIT commercialself-hosted1.1 GB93.0 88.8 90.9not scored on belebele-ces-ces, belebele-en-q-cs-doc, webfaq-ces, wiki-cs
qwen3-emb-8b Apache-2.0 commercialself-hosted15.1 GB— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
nemotron-3-embed-8b No published Czech evaluation.OpenMDW-1.1 commercialself-hosted15.9 GB— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
pplx-embed-4b Instruction-free by design (model card).MIT commercialself-hosted8 GB— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
harrier-27b Gemma-3 backbone; needs the H100 NVL (bee).MIT commercialself-hosted54 GB— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
octen-8b Apache-2.0 commercialself-hosted15.1 GB— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
granite-311m CPU tier.Apache-2.0 commercialself-hosted0.6 GB— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
e-infra-qwen3-emb-4b Hosted fallback on the e-INFRA CZ API; exact served revision unknown.Apache-2.0 commerciale-INFRA API—— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
voyage-4-large commercial API commercialAPI—— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank
gemini-embedding-001 Vertex EU; backend not implemented yet.commercial API commercialAPI—— — —not scored on belebele-ces-ces, belebele-cs-q-en-doc, belebele-en-q-cs-doc, webfaq-ces, wiki-cs, wiki-cs-rerank

* scored on a query subsample, which cannot decide a pick.

Rerankers

decided by the mean over these slices
Rerankers candidates
ModelLicenceHostingVRAMwiki-cs-rerank MAPMeanStatus
bge-reranker-v2-m3 Apache-2.0 commercialself-hosted1.1 GB96.5 *96.5only a query subsample on wiki-cs-rerank
qwen3-rr-4b Prompt format from the model card / vLLM example; served via seq-cls override.Apache-2.0 commercialself-hosted8 GB— —not scored on wiki-cs-rerank
nemotron-rr-1b vLLM needs the score chat template from the model card, else scores are silently wrong.OpenMDW-1.1 + Llama 3.2 commercialself-hosted2.5 GB— —not scored on wiki-cs-rerank
jina-rr-v3.5 research-only (non-commercial licence).CC-BY-NC-4.0 research-onlyself-hosted1.2 GB— —licence does not allow commercial use
querit-4b Optional wildcard; custom loading, not wired to a backend yet.Apache-2.0 commercialself-hosted—— —not scored on wiki-cs-rerank
e-infra-qwen3-rr-4b Hosted fallback on the e-INFRA CZ API.Apache-2.0 commerciale-INFRA API—— —not scored on wiki-cs-rerank
cohere-rerank-v3.5 commercial API commercialAPI—— —not scored on wiki-cs-rerank

* scored on a query subsample, which cannot decide a pick.

Slices

Selection slices and published reference scores
SliceKindDescriptionPublished reference
belebele-cs-q-en-docretrievalCzech questions over English passages (Belebele eng_Latn-ces_Latn)BAAI/bge-m3 nDCG@10 93.0
wiki-cs-rerankrerankingRerank 1 positive among 9 hard negatives (Czech Wikipedia)BAAI/bge-m3 MAP 88.7

Compute

Open models run as offline batch jobs on MetaCentrum (PBS Pro, Singularity, vLLM). Their outputs are imported into the request cache, so rerunning the evaluation needs no GPU access. Latency for API stacks is measured from Prague; for open models on the node that runs them.

Computational resources were provided by the e-INFRA CZ project (ID:90254), supported by the Ministry of Education, Youth and Sports of the Czech Republic.