Local AI Token Speed Ranking
Compare token speed across devices under the same model, quantization, engine, context length and concurrency. Every value carries a trust grade, and calculated or estimated numbers are never presented as measurements.
Own-measured Third-party Calculated / estimated User-submitted · evidence verified 1625 measurements · 93 devices · 132 models (886 own-measured · 733 third-party) · Updated 2026-10-30
What are trust grades?
The same number can be more or less reliable depending on how it was obtained, so every value carries one of the grades below.
- Own-measured
- Values GitSeonjeap measured itself with a consistent method and recorded in a video or log. The default ranking is based on this grade.
- Third-party
- Values published by other people or organizations, copied with their source. Their conditions may differ from ours, so they are left out of the default ranking.
- Calculated / estimated
- Not a measurement: values calculated or estimated from other measurements. A real device may differ, and they are left out of the default ranking.
- User-submitted · evidence verified
- Submitted by a visitor and published after review confirmed that the evidence image (model, speed, device info) matches the entered values. We did not measure it ourselves, so it has its own grade.
- Pending review
- Submitted but not yet reviewed. It has not been confirmed as accurate, so it is not shown in the ranking by default.
Only groups with 3 or more devices measured under the same conditions are listed. Click one to fill in all conditions at once.
Results
By default only own-measured values and evidence-verified submissions are shown. Other grades can be included with the Trust grade filter.
Could not load the data. Please try again in a moment.
-
1
NVIDIA GPU · 192 GB UD-IQ4_XS · llama.cpp · Context length 0 (unspecified)n=1 98.5–98.5 #1Own-measured The test video was taken down or is private
⚠ Has caveats
- 전력 한도 300W/장(Max-Q) - 600W 워크스테이션판(h3 프로젝트의 PRO 6000)과 다른 카드
- 2장 합계 값(peak_mem·power_w)
- dup
-
2
NVIDIA GPU · 64 GB UD-IQ4_XS · llama.cpp · Context length 0 (unspecified)n=1 72.8–72.8 74% of #1Own-measured The test video was taken down or is private
⚠ Has caveats
- 전문가를 시스템 RAM 으로 오프로딩(--n-cpu-moe)이라 VRAM 만의 속도가 아님
- 서버 A 는 -t 32 (cgroup CPU 쿼터 61코어 문제 후 스레드 고정)
- 2장 합계 값(peak_mem·power_w)
- 서버가 중간에 SSH 끊김
- 줄인 범위(jobs=server_a_5090x2_short)
- Flash-Next 는 --n-cpu-moe 8 + per_layer_token_embd CPU
-
3
DGX family · 128 GB UD-IQ4_XS · llama.cpp · Context length 0 (unspecified)n=1 29.4–29.4 30% of #1Own-measured ▶ 무료 챗GPT급 로컬 AI, 24GB 한 장이면 된다 — RTX 3090 · 5090 · 맥 미니 M6 직접 재 봤다
⚠ Has caveats
- DGX 담당과 중복 가능(대여 서버라 GPU 담당이 수집)
- power_note power.limit=N/A
- peak_mem=시스템(통합) 메모리 사용 최대
- peak_mem=시스템(통합) 메모리 최대 GiB - VRAM 아님
-
4
Mac · 32 GB UD-IQ4_XS · llama.cpp · Context length 0 (unspecified)n=3 5.2–5.7 5% of #1Own-measured ▶ GLM-5.3 Flash vs 큐웬 3.8 Flash-Next — 맥 미니 32GB · RTX PRO 6000 · 5090 직접 쟀다
⚠ Has caveats
- Flash-Next UD-IQ4_XS 93.7GB는 32GB보다 커서 CPU(-ngl 0) + 전문가 SSD 읽기
- 1회 측정(stddev 0)
- 영상 제목 "초당 5.7토큰"은 이 행(-t 8 tg32 5.71)이고 서버 실측(-t 6)은 4.25~4.60 — 조건 다름
There are no measurements for these conditions yet.
Try widening the trust grade or changing the context length or concurrency. If you have measured this yourself, you can submit it and it will be added after review.
Submit my measurementn is the number of own measurements under the same conditions. When n is 2 or more, the value shown is the median, with the min–max range beside it. When n is 1 it is a single run and may differ when measured again.