Guides / Computing costs and local AI

Our 2017 iMac couldn't run Ollama on macOS. On Linux it reached 70% of a modern desktop

In our October 2026 test, Ollama needed macOS 14; on Ubuntu, a 2017 iMac ran 3B and 8B models at about 70% of our Dell’s speed.

Engineering note

By Steven Michael Barber · Published October 21, 2026 · Updated November 3, 2026 · 5 min read

In part 1 we measured one modern desktop with no dedicated graphics card running local language models through Ollama. The longer question is whether the computers and phones we already own can work together as one private AI network. That needs a second machine, so we turned to the oldest one on our list: a 2017 21.5-inch iMac that had always felt slow.

What this part shows: the Ollama version we downloaded in October 2026 would not run on the iMac's installed macOS (13.6.7). After we replaced macOS with Ubuntu Server, it ran the same 3B and 8B models at about 70% of our Dell's speed, the 8B model fit in its 8 GB of memory, and the first load of each model took 15 to 76 seconds. At this stage we had not split a model across machines, used the iMac's graphics card or tested the phones.

The machine as we found it

We read the specifications from the machine itself before changing anything:

For comparison, part 1's Dell has a 20-core Intel Core Ultra 7 265 and 32 GB of memory.

Ollama on macOS: refused

Ollama's install script downloaded the app and then reported Unable to find application named 'Ollama'. The app's own settings explain why: it declares a minimum of macOS 14.0, and this iMac runs 13.6.7. We did not try to upgrade macOS. If you're in the same position, Apple's compatibility list for macOS 14 tells you whether your Mac can make that jump. The command-line client installed, but no server ever started, so there was nothing to benchmark. To check an older Mac yourself, defaults read /Applications/Ollama.app/Contents/Info.plist LSMinimumSystemVersion prints the requirement.

Replacing macOS with Ubuntu Server

The iMac would be a dedicated worker, so we chose Ubuntu Server 26.04 LTS, with no desktop. It keeps receiving security updates, Ollama supports it directly, and it leaves almost all of the memory for models. Our files were already backed up, so we let the installer erase the internal disk. Three lessons from the install:

• Check what a drive really is. We planned to install on an external drive we believed was an SSD. The installer showed its model number, which belongs to a line of portable hard disks, so we used the internal disk instead.
• The Wi-Fi worked without extra drivers. The chip identified itself on macOS as a Broadcom BCM43602. Linux supports it with its built-in driver, and the installer offered Wi-Fi straight away.
• Get tools from their official sites. While we looked for a tool to write the installer stick, a look-alike download page offered an unrelated browser installer. Chrome blocked it. Type the official site's address yourself.

After the first boot, the operating system used 617 MiB of memory, with 6.5 GiB available. Linux reports memory in binary units, and we quote them as printed: 1 GiB = 1.074 GB and 1 MiB = 1.049 MB. We did not record macOS's memory use before erasing it, so we can't give that comparison as a number.

The same benchmark as the Dell

We installed Ollama with its official Linux script and ran part 1's test: the same prompt, temperature 0, a 256-token limit and three runs per model. This time each test began with the model unloaded, so the first run includes loading it.

Every answer stopped at the 256-token limit. The 8B model fit: while it ran, the iMac used 5.0 GiB with 2.1 GiB still available, and swap stayed at the same 1.1 MiB throughout. The graphics card sat idle: the installer printed "AMD GPU ready", but Ollama found only the processor.

First loads were slow. The Dell's first 8B run reported 5.4 seconds of loading. We suspect the iMac's old hard disk, but we did not test that: we did not measure disk activity or clear the file cache, and we did not record the Dell's storage. The roughly 14-fold gap is an observation, not a diagnosis.

Worked example: reading the 8B cold start

The 8B model's first run took 125.79 seconds in total. Ollama attributes 75.81 seconds of that to loading the model and the remaining 49.98 seconds to reading the prompt and writing 256 tokens. At 5.31 tokens per second, the writing alone takes about 256 ÷ 5.31 = 48.2 seconds. The second and third runs, with the model already in memory, each took 48.41 seconds. So the first question costs more than two minutes, and the questions right after it cost under one. Note that Ollama unloads a model after five idle minutes by default; our service log shows OLLAMA_KEEP_ALIVE set to 5 minutes. A worker that waits longer than that between requests pays the load again, unless you raise that setting.

What 70% means

The 2017 quad-core iMac produced tokens at about 70% of the Dell's rate. One likely reason is that generating each token mostly means reading the model's weights from memory, and the two machines' memory speeds differ less than their core counts. We have not tested that, so treat it as a hypothesis. The iMac also ran Ollama 0.40.7 on Ubuntu and the Dell 0.40.4 on Windows, so software differs as well as hardware. As in part 1, this measures speed, not answer quality.

The first network number

For splitting one model across machines, delay matters. With both machines on Wi-Fi, 20 pings from the Dell to the iMac all came back. After a slow first reply of 171 ms, the other 19 took 8 to 12 ms, median 10 ms. We measured delay only, not bandwidth.

The next experiment

Part 3 used llama.cpp's experimental RPC mode to split one model across the Dell and the iMac. Every Wi-Fi split we measured generated text more slowly than the Dell alone. That follow-up compares each machine's baseline with the split and separates the measurements from our explanations for them.

The lesson so far: before giving up on old hardware, check whether the software you need still supports it. On its installed macOS, current Ollama would not run on this iMac at all. On a current Linux release, it did real work.