What runs on what

Measured on a real machine, not asserted. Nothing here is enforced: the installer accepts any computer and reports what it found. Hardware decides one thing only — how fast local models run, which is one supplier class of five. Cloud agents, tools, projects and the relay run on any machine that runs Cockpit.

Tier	Machine	What you get
1. Run Cockpit	4 GB RAM, about 10 GB disk, no GPU	Everything except local models: paid and free-tier CLI agents, free REST workers, tools, projects, the library, the relay.
2. Local models, usable	8 GB RAM, no GPU	A 3B model at about 8 tokens a second and a 1.5B model at about 16 — both above the line where answers time out.
3. Local models, comfortable	4 GB video memory + 16 GB RAM	A 3B model at about 58 tokens a second, a 1.5B at about 104. Do not expect a 7B (about 5.6) or a 14B (fails, see below).
4. Larger models	8 GB video memory for 7B-class; 12 GB or more for 14B-class	Inferred from the fit rule, not measured — the test machine has a 4 GB card. We will replace this row with numbers when we have them.

What “usable” means

About 10 tokens a second feels live: faster than you read. Below about 4 tokens a second answers do not feel slow, they time out: Cockpit's own ceilings are 180 seconds for the planner, 120 for a chunk and 60 for rules.

Surprises we measured

Core count barely matters. A 3B model on the CPU ran at 8.1 tokens a second on 2 threads and 6.8 on 12: CPU inference is limited by memory bandwidth, not cores. (One machine; other machines untested.)
The cliff is between 3B and 3.8B on a 4 GB card: 58 down to 12 tokens a second, because the model stops fitting.
A 14B model on 4 GB does not run slowly, it fails: a 500-token answer took 216 seconds against a 120–180 second ceiling.
A small GPU only pays off up to about 3B: 7× faster than the CPU at 3B, only 1.5× at 7B and 1.3× at 14B.

How the app adapts

Cockpit reads the machine and adjusts what it can. Bookshelf cuts book chapters to about 12 KB with no GPU, 20 KB at 4 GB of video memory and 40 KB above, because a chapter too large for a small card took minutes of prompt processing. When two local models would not fit in video memory together, they are run one at a time. A question about your library that lands on a model too small to answer it gets a warning, with an offer to switch.

What this page does not say

These numbers come from one test machine (a 4 GB AMD card in an ordinary desktop) and from the models we tried. They show where the limits are, not what every machine will do. For what the free and paid cloud models offer, see AI Models. For the full measurements and how to repeat them, see the Setup Guide. Download.