GitSeonjeap
English

Sign in

Reading needs no sign-in. Sign in to write posts or comments, or to recommend and report, in the community.

By signing in you agree to the notices below. Community rules · Privacy notice

Keep browsing without signing in
Only numbers we measured · with method and date

Only the numbers we measured, in one place

From local-AI token speed to AR/AI glasses — only figures that say who measured them, under what conditions, and when.

Compare AR/AI glasses → Submit your results · Coming soon

Qwen3.8-27B generation speed

Measured by us
  1. 1 RTX 5090 32GB VRAM 32GB · 전력 한도 575W(기본 600W) 76.4tok/s
  2. 2 RTX 3090 24GB VRAM 24GB · 전력 한도 280W(기본 350W) 35.9tok/s
  3. 3 DGX Spark 128GB 통합 메모리 128GB 11.8tok/s
  4. 4 맥 미니 M6 32GB 통합 메모리 32GB 8.0tok/s
Context 4K · UD-Q4_K_XL · llama.cpp b11311 · 1 concurrent request · measured 2026-10-01 Four devices measured by us on the same model, quantization, engine and context. The full ranking is still being built.
1,000+Measurements by us
10Devices measured
45Models tested
5AR/AI glasses compared

Sections

Each section compares things differently. Local AI by measured numbers; glasses by specs and real-world use.

Local AI

Local-AI token speed, memory and concurrent users. Mac, GPU and DGX setups compared under the same conditions.

Preparing the data

AR/AI glasses

Specs, reviews, what's in the box, prescription lenses and Korean-language UI for AR/AI glasses, with picks by your conditions and situation.

Live now · Open →

Community

Share local AI setups, hardware choices and AR/AI glasses buying questions with just a nickname, no sign-up.

Open · Open →

Every number carries a trust grade

Grades differ in shape as well as color, so they still read in black and white.

Measured by us

Figures we measured ourselves, with the original records kept.

Measured by others

Figures measured by other people, such as reviewers or public tests. The source is always shown.

Calculated or estimated

Things like bandwidth divided by model size. Kept visibly apart so they never pass as measurements.

Submitting your own results is coming

We can't measure every machine ourselves, so we want to accept yours with proof. It isn't open yet.

  1. MeasureRun the given command to measure token speed under the same conditions.
  2. Add proofAttach a result screen that shows the model, speed and system info, and enter the token count.
  3. Reviewed, then publishedOutliers are filtered out, and confirmed submissions go live with an 'Evidence verified' mark.

Recent updates

  • AR/AI glasses comparison is live — 14 videos, picks by situation, and values corrected by web checks
  • Korean, English and Japanese are now available