Fifteen downloadable embedding models cover English, multilingual, and code — and none strictly needs a GPU

Asked (summary):

List common sentence-embedding models for 1) English text, 2) multilingual text, 3) code — up to 5 per case, dense or dense+sparse+multivector — reporting model id, release date, vector types and dims, best MTEB scores, context length, GPU need, CPU-only throughput/RAM/AWS price, and GPU VRAM/floor card/throughput/AWS price.

The rows cover 15 of 15 selected models — five per case — chosen by cross-source web prominence and technical fit, backed by at least five independent hosts. Context length is in tokens; RAM and VRAM are rough sizing estimates from parameter count (GB). Standardized CPU/GPU tokens-per-second and AWS hourly prices were not publicly available for these exact models, so those columns are intentionally blank.

dense vector (number = dimensions) sparse output multi-vector output other (Matryoshka sub-vectors)  ·  grey bars: context length (tokens), CPU RAM and GPU VRAM estimates (GB). Hover a row for its benchmark score and GPU floor card.
BGE-M3 is the clearest hybridIt is the only model here emitting dense (1024-d), sparse, and multi-vector outputs at once, with an 8,192-token context — one model for three retrieval styles.
Long context lives in multilingualgte-multilingual-base and BGE-M3 take 8,192 tokens; every selected English model stops at 512. gte also offers Matryoshka elastic dense sub-vectors.
Small English models are CPU-friendlybge-small-en-v1.5 and e5-small-v2 are 33M-parameter models with ~2 GB estimated CPU RAM — and still score above 59 on the MTEB 56-task average.

BAAI/bge-large-en-v1.5 posts the best English MTEB 56-task average here, 64.23, ahead of bge-base (63.55) and e5-large-v2 (62.25). arxiv.org

multilingual-e5-large-instruct leads the multilingual picks on MTEB(Indic) at 70.2, on a 560M-parameter body estimated at ~5 GB CPU RAM. huggingface.co

Among code models, nomic-embed-code is by far the largest (7B parameters, ~39 GB CPU RAM estimate, ~20 GB FP16 VRAM) — the only pick needing more than a T4-class card. graysoft.dev

Code benchmarks are not comparable to MTEB: codebert-base reports 0.87 recall@10 on CodeSearchNet Java, unixcoder-base 37.33 NDCG@10 on CoIR — different metrics on different tasks. arxiv.org

All 15 models, all requested columns

CaseModelRel.DenseSparseMulti-vecOtherBest scoreCtxGPU neededCPU tpsCPU RAMCPU $/hVRAMGPU floorGPU tpsGPU $/hSource

Method: 15 selected downloadable embedding models (5 English, 5 multilingual, 5 code), chosen by cross-source mentions plus technical fit; sourced from at least five independent hosts, no URL backing more than two rows. Scores keep each source's own benchmark and metric — English MTEB averages, multilingual MTEB subsets, and code-retrieval metrics are not comparable across cases. CPU RAM / GPU VRAM are rough estimates from parameter count (FP32 or FP16 weights × 1.35 + ~1 GB, 2 GB floor), not measurements; GPU floor card is an estimated capacity class. CPU/GPU tokens-per-second and AWS hourly prices lacked standardized public figures and are left blank. Blank release dates or dims were not verified in collected sources.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.

Keenable SELECTAsk your own question
Made with Keenable SELECT