Making it fast

anz33 routes by measured speed. Until you measure, it is working from a first guess. One command fixes that, and it is the single highest-value thing you can do after installing.

Measure your machine

anz33 bench

This sends a real, short request to every model on every running engine and times the answer. Expect it to take a few minutes the first time — a model has to load into memory before it can be timed, and that loading is part of what you are measuring.

To time just one engine:

anz33 bench ollama

Reading the result

Speed is reported in tokens per second. A token is roughly three-quarters of a word, so:

Under 10 tok/s	Slower than reading aloud. Usable for short answers; painful for long ones.
20–40 tok/s	About as fast as you can read. This is the comfortable range on a modern laptop.
60+ tok/s	Faster than you can read. Typical of a small model on a good GPU or Apple Silicon.

The numbers are saved in ~/.anz33/bench.json and shown next to each model on the dashboard.

When to measure again

After installing a new engine or pulling a new model — an unmeasured model cannot be chosen confidently.
After a big OS or driver update, which can change GPU performance substantially.
If routing starts feeling wrong. Numbers from six months and two model upgrades ago are not describing your machine any more.

Battery

Unplugged, the fastest engine is often the one that empties your battery quickest. anz33 knows whether you are on battery and factors it in. To lean on that harder, turn on battery saver from the dashboard, or in ~/.anz33/config.json:

"battery_saver": true

Things that are slow for other reasons

If everything feels slow and bench disagrees, the cause is usually one of these:

First request after a pause. Engines unload models from memory when idle. The first request pays the loading cost again.
The model is too big for your machine. Once a model does not fit in memory, the system starts swapping to disk and speed collapses. A smaller model is not a compromise here — it is the fix.
Something else is using the GPU. Video calls, games and video editors compete for the same hardware.
A cloud engine is answering. Cloud latency is network latency. Check the engine column on the dashboard.

What it saved you

anz33 savings

An estimate of what the same requests would have cost at cloud prices, based on the tokens you actually put through local engines.