The machines
Server
Alienware desktop
- Graphics
- NVIDIA RTX 4090, 24 GB
- Processor
- Intel Core i9-13900F
- Memory
- 64 GB
- System
- Windows 11
Always on. Runs Ollama and serves models to every other computer in the house.
Uses the server
Mac mini M4
- Chip
- Apple M4
- Memory
- 16 GB
Runs small models itself, and uses the PC for anything bigger.
Uses the server
MacBook Pro
- Chip
- To be added
- Memory
- To be added
The day-to-day laptop. Talks to the PC over Wi-Fi at home.
How they fit together
One computer does the heavy lifting. Models live on the PC, whose graphics card reads memory at 1008 GB/s, against 120 GB/s for the Mac mini. The Macs send questions over the home network, and answers come back as if the model ran on the Mac.
Why not just use the Macs? A 16 GB Mac mini can give a model about 11 GB, which is enough for good small models. The PC’s card has 23 GB of much faster memory, so it runs bigger models, quicker.
What runs where
Every model on the site, at 4-bit with 8K tokens of conversation. 7 of them only fit on the PC.
| Model | On the PC | Tokens/s | On the Mac mini | Tokens/s |
|---|---|---|---|---|
| Llama 3.2 3B | Runs well | 188–313 | Runs well | 22–37 |
| Gemma 4 E4B | Runs well | 158–264 | Runs well | 19–31 |
| Qwen3 8B | Runs well | 81–136 | Runs well | 10–16 |
| Llama 3.1 8B | Runs well | 84–140 | Runs well | 10–17 |
| Gemma 3 12B | Runs well | 50–84 | Won't fit | — |
| gpt-oss 20B | Runs well | 225–375 | Won't fit | — |
| Mistral Small 3.2 | Runs well | 30–50 | Won't fit | — |
| Gemma 4 26B-A4B | Runs well | 182–303 | Won't fit | — |
| Qwen 3.6 27B | Runs well | 27–45 | Won't fit | — |
| Gemma 4 31B | Fits, just | 23–39 | Won't fit | — |
| Qwen 3.6 35B-A3B | Fits, just | 238–397 | Won't fit | — |
| Llama 3.3 70B | Runs slowly | 1.2–1.9 | Won't fit | — |
- Estimate: Calculated from a stated rule of thumb. Real results vary.Speeds: the site’s estimates, until measured on these machinesChecked: not yet verified
Llama 3.3 70B is the interesting case. It doesn’t fit on the card, so part of it Spilling overWhen a model is too big for the graphics card, part of it runs from the PC’s slower main memory.Read the full entry into the PC’s 64 GB of main memory. It will run, slowly. Measuring how slowly is one of the first tests.
Setting up the server
If you have a PC with a good graphics card, you can copy this. Install Ollama on the PC first, using Start here. Then:
Let Ollama answer the network
About 2 minutes
By default Ollama only answers the PC it runs on. Quit Ollama from the icon in the taskbar, then run this in PowerShell and reopen Ollama:
PowerShell setx OLLAMA_HOST "0.0.0.0:11434"You should seeSUCCESS: Specified value was saved.
If Windows asks whether to allow Ollama through the firewall, allow it on private networks only.
Find the PC’s address
About 1 minute
PowerShell ipconfigLook for IPv4 Address under your Wi-Fi or Ethernet adapter, such as
192.168.1.50. Reserve it for the PC in your router’s settings, usually under “DHCP reservation” or “fixed IP”, so it doesn’t change.Keep it awake
About 1 minute
A sleeping PC can’t answer. In Settings, System, Power, set the PC never to sleep when plugged in. The screen can still turn off.
Check from a Mac
About 1 minute
On the Mac, use the PC’s address in place of the example:
Terminal curl http://192.168.1.50:11434You should seeOllama is running
Then chat with a model on the PC from the Mac’s Terminal:
Terminal OLLAMA_HOST=192.168.1.50:11434 ollama run qwen3.6:35bApps that work with Ollama usually have a setting for its address. Set it to
http://192.168.1.50:11434.
What I’m measuring
This setup is also the site’s test bench. Each run replaces an estimate:
| Measurement | Replaces | Status |
|---|---|---|
| Writing speed for each model on the RTX 4090 | The speed estimates for every NVIDIA card | Not yet run |
| Writing speed for small models on the Mac mini | The speed estimates for every Mac | Not yet run |
| How much graphics memory Windows keeps back | The 0.8 GB reserve we assume | Not yet run |
| Speed when a 70B model spills into the PC’s main memory | Our estimate for models that spill | Not yet run |
| Extra delay when a Mac uses the PC over Wi-Fi | Nothing yet: we don’t estimate it | Not yet run |
| Power the PC draws idle and while writing | Nothing yet: we don’t estimate it | Not yet run |
Results, with the exact model, settings and tool versions, will go on How we check.