Build 04Placeholder
Multi-user family server
A few people sharing one model, queued or overlapping, without buying a GPU per person.
Headroom past a single session, a queue you understand, and idle power you have metered. Multi-user here means a household, not a datacenter.
- Price
- Placeholder
- Price pulled on
- Placeholder
- Source
- Placeholdername the sold listing or retailer page on the day it is pulled
- Idle
- Placeholder
- Load
- Placeholder
- Model
- Placeholder
- Quant
- Placeholder
- Context
- Placeholder
- Gen tok/s
- Placeholder
- Prompt tok/s
- Placeholder
Log idle watts for a machine that stays up, and load watts with one active request. A second stream gets its own note.
Measure one stream first. A second stream is a separate row, not a doubled number.
Don’t buy this if
- You are the only user. Pay for the single-user build that matches the model you run.
- You need separate accounts, logs, or quotas, and you have not picked software that does that. The chassis will not invent it.
- The listing is a loud server and you have nowhere to put the noise or the heat.
Parts
| Part | Item | Price (USD) | Source | Notes |
|---|---|---|---|---|
| Host | Placeholdera machine with memory past a single resident model | Placeholder | Placeholder | The extra RAM or VRAM is so the second person does not unload the first model. |
| Storage | PlaceholderSSD for the models the household actually uses | Placeholder | Placeholder | Fewer models resident is a feature. Write down which one stays loaded. |
| Network | Placeholderwired link, placed where the noise is tolerable | Placeholder | Placeholder | A closet server people can hear will get unplugged. |
| GPU | Placeholderoptional, and only after the CPU path is measured | Placeholder | Placeholder | One shared GPU is a queue. Say so on the page when you time it. |
No affiliate links on this page yet. If one is added, the source cell will say so. See the affiliate disclosure.
Gotchas
- Two people can share one resident model and wait their turn. That is a different result from two models loaded at once. Record which one you ran.
- Ollama and similar servers queue by default. The software name and version sit in the benchmark row next to the speed.
- A family server that idles all day is a power bill. Idle watts are part of the build, not a footnote.
“Multi-user” is an easy phrase to overbuy. For a household it usually means two or three people, a shared model, and a queue. It does not mean simultaneous full-speed generation for everyone, and it does not mean a second GPU per person. The honest build has enough memory that a second request waits on compute instead of waiting on a model reload.
The first measurement is still a single stream: model, quant, context, prompt tokens per second, generation tokens per second. The second measurement is the same run with another request in flight, written as its own benchmark row, with a note that says what the other request was. Averaging them into one heroic number would hide the wait.
Software choices belong in the row. A queue, a timeout, and a context limit are configuration, and they change what “the server can handle” means. When the version changes, the old row stays, with its date. This page then points at the newer one.
If the household only needs short chat, the cheap 8B build may be enough and quieter. If someone wants a large quant that only fits in a lot of RAM, the host is the used Xeon, and this page is about who else has to live with that machine’s fans and its idle draw.
Models that point here
A judgment from the model card and the size, not a speed ranking. Memory and the measured-or-not speeds are on each model page.