BGE-M3 for Multilingual Search: A Pre-Purchase Cost Checklist
In short: BGE-M3 is a multilingual embedding model that returns dense, sparse (lexical), and multi-vector (ColBERT-style) representations for 100-plus languages in a single forward pass, which makes it a cost-efficient retrieval backbone for Korean-and-multilingual search where you would otherwise pay for three separate models.
BGE-M3 is a multilingual embedding model that returns dense, sparse (lexical), and multi-vector (ColBERT-style) representations for 100-plus languages in a single forward pass, which makes it a cost-efficient retrieval backbone for Korean-and-multilingual search where you would otherwise pay for three separate models. The question this checklist answers is narrow: before you buy or rent GPUs to run BGE-M3, can you judge the deal on monthly cost and GPU-hours alone? The short answer is that you can, if you first pin down VRAM residency, pool headroom, and query mix against your own measured numbers rather than a vendor slide.
What did Hax measure on its own stack?#
Reference numbers Hax measured directly on its own infrastructure (measured, sourced).
| Dataset item | Measured value | Date | Source |
|---|---|---|---|
| 최대 VRAM 상주(스냅샷) | 84.8 GB | 2026-07-04 | bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측) |
| 카드당 총 VRAM | 95.6 GB | 2026-07-04 | bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측) |
| 최대 GPU 사용률 | 95 % | 2026-07-04 | bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측) |
- 표본
- 3 measured metrics (Hax /data curated)
- 측정 환경
- bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측)
- 수집일
- 2026-07-04
How can you reproduce these numbers?#
Follow the source column above and our open dataset at /data.
| Metric | Hax measured | Basis |
|---|---|---|
| Peak VRAM residency | 84.8 GB (measured 2026-07-04) | bc_comfy_gpus probe |
| Min free VRAM (pool low) | 10.2 GB (measured 2026-07-04) | bc_comfy_gpus probe |
| Total VRAM per card | 95.6 GB (measured 2026-07-04) | bc_comfy_gpus probe |
| GPU cards | 4 (measured 2026-07-04) | bc_comfy_gpus probe |
| Peak GPU utilization | 95% (measured 2026-07-04) | bc_comfy_gpus probe |
| Stored memories | 9871 (measured 2026-07-11) | curator stats |
| Active memories | 9577 (measured 2026-07-11) | curator stats |
| Avg confidence | 0.6810 (measured 2026-07-11) | curator stats |
| AI crawler hits 7d/6 bots | 1317 (measured 2026-07-11) | telemetry/funnel |
| Monthly GPU cost | not measured / 측정대기 | estimated only; see below |
Note: curator stats are measured 2026-07-11; the GPU snapshot is measured 2026-07-04. Cost figures below are estimated (추정) because Hax does not currently meter a per-model dollar line.
Hardware checklist. First, size VRAM against your real corpus. On our four-card pool, peak residency reached 84.8 GB with a pool low of only 10.2 GB free against 95.6 GB per card (all measured 2026-07-04). That thin 10.2 GB margin is the number that decides whether BGE-M3's multi-vector mode fits alongside your existing workloads, or whether you queue and stall. Second, watch utilization: our peak hit 95% (measured), so a BGE-M3 batch job lands on already-hot cards, not idle ones. Third, count cards honestly (four, measured) before assuming linear throughput.
Software and recall checklist. BGE-M3's value for Korean is that dense and sparse signals are produced together, so lexical Korean tokens and semantic meaning are both scored without a second model. Validate recall on your own index, not a public benchmark: our curator holds 9871 stored and 9577 active memories at an average confidence of 0.6810 (all measured 2026-07-11), and confidence at that level is where hybrid dense-plus-sparse reranking earns its keep. If your active-to-stored ratio and confidence look like ours, BGE-M3 hybrid retrieval is a reasonable buy.
Cost judgment. With no metered dollar line, treat monthly cost as estimated (추정): monthly_cost ≈ GPU_hours × hourly_rate, and GPU_hours scale with peak utilization (95%, measured) and residency headroom (10.2 GB low, measured). Rent short before you buy, measure your own residency, then decide.
Internal evidence: Hax data.
Related reading: linktest, probe
Related reading: 임베딩·시맨틱 검색 모델, 2026 현황과 추천, 로컬 RAG 문서 질의응답, 5분 시작 가이드
Responses
No responses yet. Be the first to respond.