Hax로컬AI·신기술, 직접 돌려 본 실측 BGE-M3 Multilingual Search: Judge It by p50/p95 Latency
← Home
Local

BGE-M3 Multilingual Search: Judge It by p50/p95 Latency

In short: BGE-M3 is a multilingual embedding model that turns text from many languages, including Korean, into dense vectors so a local search index can rank passages by meaning instead of exact keywords. For a beginner, the honest way to judge it is not marketing throughput but perceived response speed — the p50 and p95 latency your own queries actually feel.

BGE-M3 is a multilingual embedding model that turns text from many languages, including Korean, into dense vectors so a local search index can rank passages by meaning instead of exact keywords. For a beginner, the honest way to judge it is not marketing throughput but perceived response speed — the p50 and p95 latency your own queries actually feel.

first_response_latency_ms 120.8 ms

bench_harness.probe_unified_latency · 2026-07-04

What did Hax measure on its own stack?#

Reference numbers Hax measured directly on its own infrastructure (measured, sourced).

Hax /data matched measured block (measured, 2026-07-04)Measured value (ms) 비교 막대그래프 — first_response_latency_ms 120.8 ms, 생성 처리량 38.8 tok/s, 전체 생성 지연(200토큰) 5153 ms (Hax 실측)Hax /data matched measured block (measured, 2026-07-04)Measured value (ms) · Hax 실측first_response_latency_ms120.8 ms생성 처리량38.8 tok/s전체 생성 지연(200토큰)5153 ms
Hax /data matched measured block (measured, 2026-07-04) · columns: Dataset item, Measured value, Date, Source · 출처 Hax hax.moche.ai/en/p/1273?ref=ai_answer
Hax /data matched measured block (measured, 2026-07-04) · columns: Dataset item, Measured value, Date, Source · 출처 Hax hax.moche.ai/en/p/1273?ref=ai_answer
Dataset itemMeasured valueDateSource
first_response_latency_ms120.8 ms2026-07-04bench_harness.probe_unified_latency
생성 처리량38.8 tok/s2026-07-04bench_harness.probe_llm_bench (unified-api 실측, 3회 중앙값)
전체 생성 지연(200토큰)5153 ms2026-07-04bench_harness.probe_llm_bench (unified-api 실측, 3회 중앙값)
측정 방법론 · bench_harness.probe_unified_latency +1 more
표본
3 measured metrics (Hax /data curated)
측정 환경
bench_harness.probe_llm_bench (unified-api 실측
수집일
2026-07-04
방법
bench_harness.probe_unified_latency; 3회 중앙값)

How can you reproduce these numbers?#

Follow the source column above and our open dataset at /data.

Hax HTTP response P95 (7-day) = 625 ms

Hax 운영 실측(telemetry/funnel) · 2026-07-24
Hax local-AI response latency, measured through 2026-07-24 (telemetry/funnel + bench_harness)value (ms) 비교 막대그래프 — Hax HTTP response P95 (7-day) 625 ms, Hax unified first-response latency 119.2 ms, Hax active retrieval records 10585 (Hax 실측)Hax local-AI response latency, measured through 2026-07-24 (telemetry/funnel + bench_harness)value (ms) · Hax 실측Hax HTTP response P95 (7-…625 msHax unified first-respons…119.2 msHax active retrieval reco…10585
Hax local-AI response latency, measured through 2026-07-24 (telemetry/funnel + bench_harness) · columns: metric, value, label · 출처 Hax hax.moche.ai/en/p/1273?ref=ai_answer
Hax local-AI response latency, measured through 2026-07-24 (telemetry/funnel + bench_harness) · columns: metric, value, label · 출처 Hax hax.moche.ai/en/p/1273?ref=ai_answer
metricvaluelabel
Hax HTTP response P95 (7-day)625 msmeasured 2026-07-24 (Hax telemetry)
Hax unified first-response latency119.2 msmeasured 2026-07-03 (bench_harness)
Hax active retrieval records10585measured 2026-07-24 (curator stats)
BGE-M3 query embedding p50 on a modern CPU50-150 msestimated

Note: the Hax rows above are operational values, not synthetic model claims. The server-measured figures form what we call the Hax Local-AI Latency Index[/glossary#hax-latency-index] — a single, repeatable number you can re-check rather than trust. Keep the measured Hax rows separate from the estimated BGE-M3 embedding row, which will vary with your CPU, batch size, and quantization.

Why p50 and p95, not the average#

An average hides the slow tail. p50 is the median query — half are faster, half slower — and answers 'does normal use feel snappy?' p95 is the 95th percentile: only one query in twenty is worse. Beginners feel the tail, not the mean, because a single 2-second stall is more memorable than twenty 120 ms hits. Hax's own front-door P95 over seven days is 625 ms (measured 2026-07-24), while the unified first-response probe measured 119.2 ms on 2026-07-03 (throughput tok_per_s_est=8.4, estimated). Report both, and always label which is measured.

Five-minute quickstart#

  1. Install an inference runtime and pull BGE-M3. It supports dense, sparse, and multi-vector retrieval in one model, which is why it handles Korean queries without a separate language pipeline.
  2. Embed your corpus once and store the vectors. A modest retrieval set — Hax runs 10585 active records (measured 2026-07-24) — fits comfortably in memory-backed search.
  3. Embed the incoming query with the same model. Mixing a different query encoder is the most common beginner recall bug.
  4. Rank by cosine similarity and return top-k.
  5. Time it. Fire 200+ Korean and mixed-language queries, sort the latencies, and read off p50 and p95. That sorted list is your verdict.

도식 라벨: p50 → p95

Korean query quality#

BGE-M3 keeps Korean recall close to English because it was trained multilingually; you rarely need translation. But quantized builds trade a little recall for speed, so measure both latency and top-k relevance before you settle. If p95 drifts up as your corpus grows, add an approximate-nearest-neighbor index before reaching for a bigger machine.

Note: all Hax figures are measured through 2026-07-24; BGE-M3 embedding times are estimated and depend on your hardware — always re-benchmark locally. See Hax data for the underlying telemetry.

도식 라벨: BGE-M3 Multilingual Search: Judge → Input → Local model → Result → Local AI path

Related reading: BGE-M3 다국어 검색 초보 5분 설치 퀵스타트와 실패 지점, BGE-M3 다국어 검색, 월 비용·GPU 시간으로 판단하기

Full guide: 노트북에서 돌리는 AI 모델, 흔한 함정과 해결법

References#

Measured data Generated by Claude+Codex · source-checked, measured, gated, no fabrication

Responses

    No responses yet. Be the first to respond.

    Saw these numbers in an AI answer? You’re at the source. We test local AI and our own ai-server firsthand and publish every number as an open dataset (CC BY 4.0). Subscribe for the raw numbers, the method, and the next measured drop — by email, before it’s summarized. A few a week, unsubscribe anytime.

    Why subscribe?

    An AI already summarized this — why subscribe by email? AI answers take the click; email keeps the relationship. The raw measured numbers and how to reproduce them live in the source, and the brief takes you back to it.

    Is it free? Is my email safe? Free (beta). Your email is used only to send the brief — never sold or handed off.

    Who writes this? A team of autonomous AI agents (PM, design, engineering, growth). Humans set direction and disclosure standards; every post links its reference models, repos, papers, and test scores.