CPU-only on a used Xeon: what actually runs
First Ollama run on Frankenstein: Qwen3 Q4_K_M on the dual Xeon and the Tesla P40, 18 Nov 2026.
A lab notebook
Adam Greene writes down what his own machines do when they run a model: tokens per second, watts at the wall, and what the parts cost on the day he checked. If a figure did not come from a run or a dated price pull, the page says PLACEHOLDER.
Measured18 measured runs in the table.
Generation speed and prompt speed, with the model, the quant, the context, and the software version. An empty cell stays empty until a run fills it.
A price names the day it was pulled and the source. Sold listings and asking prices do not share a cell. Nothing here is a live quote.
Every build says when it is the wrong machine. That note is part of the page, next to the parts, not a disclaimer in the footer.
The full version is the methodology. Build-page prices are still unpulled. The first measured speeds are in the benchmark table.
Six jobs a homelab box might be for. Prices on those pages are still placeholders. Frankenstein’s first measured speeds are in the benchmark table.
Frankenstein, 2026-11-18. Two Xeon E5-2697 v2 (24 cores / 48 threads), 288 GB DDR3 RDIMM at 1067 MT/s, Tesla P40 24 GB, Ollama 0.35.0. All rows are Q4_K_M. Generation tok/s is the median of three runs (128 tokens, temperature 0, seed 42, think off). Prompt tok/s is the first uncached prefill. n/a means that pass was already cached. Cached repeats are not listed.
| Machine | Model | Gen tok/s | Date |
|---|---|---|---|
| Frankenstein | qwen3:4b-instruct-2507-q4_K_M | 12.36 | Nov 18, 2026 |
| Frankenstein | qwen3:4b-instruct-2507-q4_K_M | 6.62 | Nov 18, 2026 |
| Frankenstein | qwen3:8b | 6.91 | Nov 18, 2026 |
| Frankenstein | qwen3:8b | 4.66 | Nov 18, 2026 |
| Frankenstein | qwen3:14b | 4.41 | Nov 18, 2026 |
| Frankenstein | qwen3:14b | 3.02 | Nov 18, 2026 |
| Frankenstein | qwen3:14b | 4.24 | Nov 18, 2026 |
| Frankenstein | qwen3:14b | 1.33 | Nov 18, 2026 |
| Frankenstein | frank-qwen3-30b-64k | 9.58 | Nov 18, 2026 |
| Frankenstein | frank-qwen3-30b-64k | 8.73 | Nov 18, 2026 |
| Frankenstein | qwen3:4b-instruct-2507-q4_K_M | 69.57 | Nov 18, 2026 |
| Frankenstein | qwen3:4b-instruct-2507-q4_K_M | 60.03 | Nov 18, 2026 |
| Frankenstein | qwen3:8b | 44.18 | Nov 18, 2026 |
| Frankenstein | qwen3:8b | 40.22 | Nov 18, 2026 |
| Frankenstein | qwen3:14b | 25.42 | Nov 18, 2026 |
| Frankenstein | qwen3:14b | 21.85 | Nov 18, 2026 |
| Frankenstein | frank-qwen3-30b-64k | 65.29 | Nov 18, 2026 |
| Frankenstein | frank-qwen3-30b-64k | 57.64 | Nov 18, 2026 |
What the model cards actually say, plus computed memory at Q4_K_M and Q8. Speeds stay “not yet measured” unless a Frankenstein row exists. What can my rig run? ranks that list against a machine you type in. The ranking stays in the browser.
First Ollama run on Frankenstein: Qwen3 Q4_K_M on the dual Xeon and the Tesla P40, 18 Nov 2026.