MoonWork Lab measures how AI actually performs on ordinary hardware.
Go looking for local LLM benchmarks and most of what you find came off an RTX 4090, a pair of 3090s, or an M2 Ultra. Great machines. Most people don’t have one.
I use a PC with a single GPU I bought for gaming. So every number on this site comes off hardware like that.
Why you can trust these numbers
- I measure hardware I own. Nothing here is copied from someone else’s benchmark.
- I publish the raw figures — not just averages, but deviation and test conditions.
- I publish the commands. You can copy them and reproduce the results on your own machine.
- I publish results that don’t flatter me. When the card I bought loses, I say it lost.
- I say when I didn’t measure something. Estimates and measurements don’t get mixed together.
That’s the whole policy. It isn’t a lofty one, but it’s followed less often than you’d hope.
What this site covers
- Real local LLM performance — which models run at what speed on which GPUs
- What tight VRAM actually feels like — fitting large models into small cards, and what it costs
- Using AI for real work — where to draw the line between cloud and local
- Occasionally, hardware developers use — only things I use every day
Who this is for
- Anyone deciding whether to buy a new GPU
- Anyone wondering if the card they already own is still good enough
- Anyone trying to cut API bills by running models locally
Who’s behind it
MoonWork — a solo developer with a day job, building local-first AI tools with what’s left of the evening. A browser extension, an offline voice assistant, a handful of self-hosted services. Everything here is built and measured by one person. No team, no sponsors, no review units.
How this site makes money
Plainly: through ads and affiliate links. What won’t happen is conclusions bending to fit those links — that’s exactly why “results that don’t flatter me” is on the list above. Any post containing affiliate links says so at the top.
I’d like to see your numbers too. If you run the same benchmarks on your GPU, let me know — especially if your VRAM is tight.