The hardware roster · one machine
the server
whole machine·13 figures, from 1 exhibit · named in 1 exhibit
This roster is in beta and still being tested: every figure is proved against the page it came from, and the way they are read and shown is still changing.
a CPU-only server: 128 GB of ECC memory, a server-class processor, and a GPU present in the box and deliberately parked
This shelf has never named the card inside it. That is a fact about what was published, not a gap in this table.
Narrow these 13 figuresevery modelmodel: gemma4:12b · 2 of 13 figuresmodel: gemma4:26b · 2 of 13 figuresmodel: Mistral Small 3.2 · 2 of 13 figuresmodel: nemotron-3.5-lightning:30b-a3b · 2 of 13 figuresmodel: phi4:14b · 2 of 13 figuresmodel: qwen3.5:9b · 2 of 13 figuresmodel: not stated · 1 of 13 figures
The rail is CSS and nothing else — no script runs on this page. It narrows by model and by power cap independently, and the line above says what each one is set to; the size of the two together is the one number it cannot count. A group or a band with nothing left in it goes with its rows, so what you are left looking at is what matched. A narrowed view is not part of the address, so it cannot be linked — each card’s own page is a URL, and that is the thing to share.
Language — decode, prefill and first token
Grouped by model and by what was measured, and inside a group sorted by power cap ascending — alphabetical and numeric, never by value. The heading of a group carries everything that does not change inside it; each row carries the value and the conditions that do. A condition the page does not state reads not stated, which is a third thing and not an uncapped card. A † means the same reading is printed in a second place on its exhibit — one reading, two printings, not two figures.
| value | what moves inside this group |
|---|---|
| gemma4:12b decode, median of five scored runstok/sQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 11.0 tok/s | |
from Two Hours at 12 tok/s § The one table to rememberread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-one-table-to-remember | |
| gemma4:12b time to first token, median of five cold-prefix runssQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · 536 prompt tokens, its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 6.9 s | |
from Two Hours at 12 tok/s § The wait: prefill is the real CPU taxread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-wait-prefill-is-the-real-cpu-tax | |
| gemma4:26b decode, median of five scored runstok/sQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 24.7 tok/s | |
from Two Hours at 12 tok/s § The one table to rememberread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-one-table-to-remember | |
| gemma4:26b the page says: gemma4:26b (MoE) time to first token, median of five cold-prefix runssQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · 536 prompt tokens, its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 3.7 s | |
from Two Hours at 12 tok/s § The wait: prefill is the real CPU taxread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-wait-prefill-is-the-real-cpu-tax | |
| Mistral Small 3.2 the page says: mistral-small3.2:24b decode, median of five scored runstok/sQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 7.1 tok/s | |
from Two Hours at 12 tok/s § The one table to rememberread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-one-table-to-remember | |
| Mistral Small 3.2 the page says: mistral-small3.2:24b time to first token, median of five cold-prefix runssQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · 1,023 prompt tokens, its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 24.6 s | |
from Two Hours at 12 tok/s § The wait: prefill is the real CPU taxread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-wait-prefill-is-the-real-cpu-tax | |
| nemotron-3.5-lightning:30b-a3b decode, median of five scored runstok/sQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 31.5 tok/s | |
from Two Hours at 12 tok/s § The one table to rememberread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-one-table-to-remember | |
| nemotron-3.5-lightning:30b-a3b the page says: nemotron-3.5-lightning:30b-a3b (MoE) time to first token, median of five cold-prefix runssQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · 532 prompt tokens, its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 4.0 s | |
from Two Hours at 12 tok/s § The wait: prefill is the real CPU taxread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-wait-prefill-is-the-real-cpu-tax | |
| phi4:14b decode, median of five scored runstok/sQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 11.1 tok/s | |
from Two Hours at 12 tok/s § The one table to rememberread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-one-table-to-remember | |
| phi4:14b time to first token, median of five cold-prefix runssQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · 505 prompt tokens, its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 8.1 s | |
from Two Hours at 12 tok/s § The wait: prefill is the real CPU taxread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-wait-prefill-is-the-real-cpu-tax | |
| qwen3.5:9b decode, median of five scored runstok/sQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 13.6 tok/s | |
from Two Hours at 12 tok/s § The one table to rememberread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-one-table-to-remember | |
| qwen3.5:9b time to first token, median of five cold-prefix runssQ4_K_M · Ollama · 0.32.15 · a 4,096-token context · 537 prompt tokens, its 32 physical cores · 2026-08-25 (UTC) · power cap and concurrency not stated | |
| 4.9 s | |
from Two Hours at 12 tok/s § The wait: prefill is the real CPU taxread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-hours-on-battery · section id #the-wait-prefill-is-the-real-cpu-tax | |
Power, temperature, and what it costs
1 row — same rule as above.
| value | what moves inside this group |
|---|---|
| not stated memory bandwidth, best triadGB/sa STREAM triad, gcc -O3 -march=native -fopenmp · version not stated · a 2,289 MiB working set, best of 9 timed reps · 2026-08-25 (UTC), 17:57–19:46Z · quantisation, power cap, context and concurrency not stated | |
| 100.7 GB/s | |
from Two Hours at 12 tok/s § Read this before the table: CPU speed is a memory-bandwidth storyread off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-hours-on-battery · section id #read-this-before-the-table-cpu-speed-is-a-memory-bandwidth | |
Where it appears
Every released exhibit that names this identity, in the shelf’s own order.
- Two Hours at 12 tok/s — 13 figures on this page
How this one is named here
The canonical name above is the full product name. Everything below is a spelling a published page or kit has actually used for it, kept so a reader who searches an older page’s words lands on this row. The set is closed and historical: no page cut after 2026-09-17 writes a memory class where it can write a card.
- the server — 128G of ECC RAM, a server-class processor.
- a server with 128G of ECC RAM and a server-class processor: 64 logical CPUs, 32 physical, not a hybrid part; a GPU present in the box and parked
- cpu-server
- the server (128G ECC)