On this pageWhy online speed charts are not enoughWrite down your setupMeasure four thingsUse the same test each timeChoose the winnerKeep a simple results sheetCan I use a creator benchmark as my result?Sources
DT

Written and reviewed by DeviceterraPractical guidance · Updated October 7, 2026

THE SHORT ANSWER

Speed numbers alone are not enough. A good test also checks whether the model completes your real task correctly.

TRY IT YOURSELF

Run a repeatable local benchmark

Online speed charts are useful clues, but they cannot tell you exactly how a model will behave on your computer. A fair local test records the full setup and uses the same prompts every time.

  1. 1

    Record the computer

    Write the operating system, CPU, RAM, GPU, VRAM, and free storage.

  2. 2

    Record the software

    Write the engine version, full model name, quantization, and context setting.

  3. 3

    Prepare ten prompts

    Use prompts from your real work. Save them in a text file so they never change between tests.

  4. 4

    Run one warm-up prompt

    The first answer may include model loading time. Run it once before measuring normal responses.

  5. 5

    Run and score the test

    Record first-word waiting time, total answer time, memory use, correctness, and instruction following.

How to know it worked
  • Someone else could repeat the test from your notes.
  • You measured answer quality as well as speed.
  • You can compare a future model against the saved result.
THE FULL GUIDE

Here is what matters, why it matters and how to check it on your own setup.

01

Why online speed charts are not enough

Model speed changes with the computer, file type, AI app, settings, prompt length, drivers, and cooling. A result from another computer is only a clue.

Test the exact model file on the exact computer you plan to use.

02

Write down your setup

  • Computer and operating system.
  • CPU, RAM, GPU, and VRAM.
  • AI app and version.
  • Exact model name, file, and quantization.
  • Context size and other important settings.
  • Whether the model was loading for the first time.
03

Measure four things

First, measure how long it takes before the first word appears. Second, measure how fast the rest of the answer arrives. Third, watch the highest memory use. Fourth, score the answer itself.

A fast wrong answer is not a good result. A correct answer that takes too long may also be useless for the job.

04

Use the same test each time

  • Create 20 real questions.
  • Use the same prompt and settings.
  • Keep the expected answer or scoring rule.
  • Include difficult and should-refuse cases.
  • Record errors, made-up facts, and broken formats.
05

Choose the winner

Pick the smallest model that passes your quality mark at an acceptable speed. Save the results so a new model must prove it is better.

Test again after changing the model, file, AI app, prompt, context, driver, or hardware.

06

Keep a simple results sheet

Make one row for each question. Record the model file, settings, time before the answer starts, total waiting time and whether the answer passed. Keep first-load time separate from later requests. Otherwise, you may blame a model for a delay caused by loading its file.

Use the same prompts for every candidate. Repeat a few requests to see whether results vary. Test with your usual browser and work apps open. Write down memory use and any crash or freeze. Do not report a best run as though it were the normal result.

A mistake to avoid

Tokens per second measures output speed, not accuracy. A model can be fast and wrong. It can also have a long first wait followed by fast writing.

07

Can I use a creator benchmark as my result?

Use it to choose what to test. Record it as someone else's result. Your hardware, file, settings and workload must match before you can make a useful comparison.

PLAIN ENGLISH

Words you may see in this guide

RAM
The working memory used by your computer and apps.
VRAM
The memory on a dedicated graphics card. Shared graphics may use system RAM.
Context
The text and other input the model can use for its current answer.
Quantization
A way to make model weights smaller. It can change memory use and answer quality.
Engine
The software that loads the model and runs it.
RAG
A workflow that finds useful document passages before asking a model to answer.
Benchmark
A repeatable test. A result is useful only when you know the setup and task.
RESEARCH SOURCES

Official facts and real user evidence

Official documentation supports product and model facts. Community discussions show real setups, failures, and questions. A community result is supporting evidence, not a promise that another computer will perform the same way.

YOUR NEXT STEP

Watch a real test. Check what fits your computer.

See practical local AI tests from DeviceTerra, then use MamiLens to build a hardware-aware shortlist for your own setup.

This guide is educational. Model software, licenses, and hardware support can change. Check official sources before an important deployment.