About: who writes this, on what machines, and how we measure

Last reviewed: 2026-10-22

freedzhost.cloud is a lab. We break things, we measure them, and we write down what happened.

The machines

Every figure on this site comes from our own hardware. No other machine is quoted as ours.

• Geekom A7: our main mini-PC.
• NUC 7 i5: a slim, old mini-PC that runs all our services. Proof that a homelab does not need new hardware.
• Gamer machine: an RTX 5080 and an external RTX 4070 Ti SUPER. This is where the local models run.
• Laptop and servers: an EliteBook workstation, plus mindhub and sentinelle for the services around the models.
• Test machines: an older-generation Dell laptop and a second NUC 7, kept for trials, experiments and things we expect to break.

How we measure

We measure tokens per second on the real cards, with the real server and the real model file, and we say which ones. We do not extrapolate from another card. For choosing a model we use fh-bench: code judges each answer and the models then work for real (the bench). GPU and memory readings come from fh-probes.

The protocol, in short

• Speed (tokens per second): the figure the llama.cpp server itself reports for each request (predicted_per_second for generation, prompt_per_second for reading the prompt). We do not time it from outside.
• Quality (fh-bench): temperature 0, and every answer is judged by code, never by another model: a tool really called, a JSON that validates, an answer that matches, a program whose tests pass. The long-context test hides a needle in a generated text at growing sizes (4k up to 128k tokens), always the same text.
• Repetitions: unless an article says otherwise, a figure is one run on one machine, not an average. Where we repeated a measurement, the article says how many times and reports the median and the range.
• Versions and dates: models, quantisations and servers change every few weeks. The review date at the top of each article is the last time we reviewed the text, not necessarily the date of each measurement; a figure is only valid for the hardware and the model file named next to it.
• Outside figures (a paper, a vendor, another team): always labelled as such, with the source linked. If it is not labelled, it is ours.

When something is not measured, the tutorial says so.

What we do not promise

• No dressed-up synthetic benchmarks: a number that only exists in a test harness is not a result.
• No income promises: running this at home saves money on subscriptions, it does not make any.
• Versions move. Each tutorial shows when we last reviewed it, and what we tested it with when that matters. Read that date before you copy a command.

Questions or corrections: write to us.

Found this useful? Tell us what to measure next.