StudyMarkets · Local AI

Real setups

Machines people actually built to run LLMs at home — the hardware they used, what it draws from the wall, and the tokens per second. Photos and numbers, not theory.

Curated: every entry is checked before it goes up. Measured numbers are labelled as measured; estimates are labelled as estimates. The two are never mixed.

Bial — three machines, 84 GB of VRAM

Brazil · Ubuntu · running models for experimentation and as a fallback for his own AI products. In his words: "I don't want to depend on anyone else's API."

Open frame with two graphics cards mounted vertically, power supply and cabling visible — Open frame, cards mounted vertically

Two assembled PC cases side by side — Two of the three machines

A third closed PC case — The third build

Measurements pending. The hardware below is confirmed by the owner. The tokens per second are this site's formula applied to that hardware — not measured on the machines. They are marked estimate and will be replaced by real numbers as soon as he runs them.

Machine 1 — dual Xeon

3× RTX 3060 12 GB · dual Xeon · 64 GB system RAM · Ubuntu
36 GB VRAM · 360 GB/s · ~660 W

Model	Fits	tok/s est
32B dense	yes	13.9
27B dense	yes	16.5
MoE 30B-A3B	yes	148.5
70B dense	no	—

Open this machine in the calculator →

Machine 2 — the white one

4× RTX 3060 12 GB · Intel i5 · 16 GB system RAM
48 GB VRAM · 360 GB/s · ~830 W

Model	Fits	tok/s est
32B dense	yes	13.9
MoE 35B-A3B	yes	148.5
70B dense	tight	6.4

Note: 16 GB of system RAM for 48 GB of VRAM is tight. The model is staged through system memory before it reaches the cards, so loading takes longer than it needs to. Going to 32 GB should cut load time — it will not change speed afterwards.

Open this machine in the calculator →

Machine 3 — AMD, still being built

4× Radeon RX 6700 XT 12 GB (planned) · in assembly
48 GB VRAM · 384 GB/s · ~1070 W

Model	Fits	tok/s est
32B dense	yes	14.8
MoE 30B-A3B	yes	158.4
70B dense	tight	6.8

When this one runs, it will be the first AMD data on this site. Every measured number we have today comes from NVIDIA or Apple hardware.

What this build shows. Stacking identical cards adds memory but not bandwidth — all three machines sit at the speed of a single card. That is why the dense 70B crawls at ~6 tok/s while a MoE model on the same hardware runs more than ten times faster: MoE needs the memory but only reads a fraction of it per token.

If you have cards sitting idle, look for MoE models first.

Send your setup

Built something? Your numbers are worth more than any benchmark we can run.

Send these, and photos if you have them:

Cards — model and how many
Motherboard, CPU, system RAM, power supply
OS and the program you use — llama.cpp, Ollama, LM Studio. This matters as much as the hardware: the same model on the same machine can differ 5× between them.
The numbers — model, quantisation, context length and tokens per second

Where to find tok/s: llama.cpp prints it at the end of each generation (eval time … tokens per second). Ollama: run with --verbose. LM Studio shows it under the response.

Please don't estimate. If you haven't measured it, we would rather publish the build without a number than publish a wrong one. Separating measured from estimated is the whole reason this site is worth reading.

Multi-card builds and repurposed mining rigs are especially welcome — almost nobody publishes those numbers.