About naxiv

naxiv exists to answer one question honestly: when you run AI yourself, on your own machine or a rented cloud GPU, what actually works, and what's worth paying for? Vendor spec sheets won't tell you real tokens/sec, and forum anecdotes are inconsistent, so we test it ourselves: the hardware, the models, and the services.

How we test

We buy or rent the gear, run the models, and use the services first-hand before we write a word. Where we publish a performance number, it comes from a real run on real hardware (same model, same quantization, same runtime) captured by our bench_runner.py harness, which wraps llama-bench -o json. If we haven't measured or used something, it doesn't get a recommendation here: no estimates, no copied spec sheets.

Who writes it

Embedded, AI and robotics engineer. Tests local LLMs and AI tools first-hand on real hardware and cloud GPUs, covering the hardware, the models, and the services.

Find Pedro Santos on github.com , linkedin.com .

How we make money

Some links on this site are affiliate links: if you buy through them we may earn a commission at no extra cost to you. This never changes our verdict; the testing comes first, the links come after. We only recommend gear and services we have tested or would use ourselves.