Skip to content
AIPCs

Which computer can run AI, and how fast?

We run real models on real machines, publish every number, and recommend computers from the results.

Our desktop, with and without its graphics card Writing speed in tokens per second (higher is better). About 10 reads comfortably.
  • RTX 4060 8 GB
  • Ryzen 9 7900X alone

* Model larger than the 8 GB card; part of it runs from system RAM.

Our desktop, with and without its graphics card, tokens per second
ModelRTX 4060 8 GBRyzen 9 7900X alone
Gemma 4 E4B67.717.8
Llama 2 7B65.515.5
Qwen3.5 9B45.510.4
gpt-oss 20B39.521.8

Tests T0002 and T0001, llama.cpp build 11191.

What our tests show so far

45.5 tokens/s
An 8 GB RTX 4060 writes with Qwen3.5 9B at several times reading speed. For models up to about 9 billion parameters, an entry-level graphics card is plenty.
21.8 tokens/s
gpt-oss 20B on a Ryzen 9 7900X with no graphics card at all. Mixture-of-experts models read only a slice of their weights per word.
16x
Less conversation memory per token for Qwen3.6 35B-A3B than for Llama 3.3 70B. Newer models handle long chats on far smaller machines.

Latest

How we work

Products are marked Tested only when we ran them ourselves, with a test ID you can check. Everything else is Researched and cites its sources. We earn commissions from some store links; they never decide what we recommend, and we link to stores that pay us nothing.

How we test How we make money Download our data