Installing Ollama on Fedora: the package split, the local service, and measured model download speed

Unlike k3s, Ollama is packaged in the official Fedora repositories:

dnf list --available ollama

ollama.x86_64   0.24.0-4.fc45   fedora

But read the packaging before installing — Fedora splits one Ollama into four packages.

Four packages, and ollama itself is 18.7 KiB

rpm -qR ollama | grep -v rpmlib

/bin/sh
ollama-base(x86-64) = 0.24.0-4.fc45
ollama-rocm(x86-64) = 0.24.0-4.fc45
ollama-vulkan(x86-64) = 0.24.0-4.fc45

ollama is a meta package: dnf install ollama drags in all three backends. Measured sizes:

Package	Installed size	Contents
ollama	18.7 KiB	README and dependency declarations only
ollama-base	44.7 MiB	main binary plus CPU backend
ollama-vulkan	49.7 MiB	Vulkan backend (generic GPU)
ollama-rocm	993.4 MiB	AMD ROCm backend

What the meta package actually installs:

rpm -ql ollama

/usr/share/doc/ollama
/usr/share/doc/ollama/README.md

Not a single binary — the real ollama executable lives in ollama-base.

Caution

ollama-rocm alone is 993 MiB. On a machine with no AMD GPU that gigabyte is wasted. It is the hidden cost of dnf install ollama.

CPU-only: install ollama-base

On a server without a GPU (like this lab box):

sudo dnf install ollama-base

That skips 993 MiB of rocm and 49.7 MiB of vulkan, keeping only the 44.7 MiB CPU backend.

Machines with an NVIDIA card cannot take this route — there is no cuda backend subpackage, so use the upstream install script that ships CUDA.

Starting the service

sudo systemctl enable --now ollama
systemctl is-active ollama
ollama --version

active
ollama version is 0.24.0

By default the service binds to loopback:

ss -lntp | grep 11434

LISTEN 0  4096  127.0.0.1:11434  0.0.0.0:*

127.0.0.1 means local access only. To serve other machines on the LAN, change OLLAMA_HOST:

sudo systemctl edit ollama

[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"

Then sudo systemctl daemon-reload && sudo systemctl restart ollama.

Caution

Binding 0.0.0.0 exposes an unauthenticated inference API to the whole subnet — Ollama has no built-in auth. If you must, put an authenticating reverse proxy in front and restrict sources in firewalld.

Pulling models: the real bottleneck

Installing the package takes seconds; pulling models is what hurts. Measured with a 91 MB model:

ollama pull smollm:135m

pulling eb2c714d40d4:  13% ▕██                ▏  12 MB/ 91 MB  154 KB/s   8m33s
pulling eb2c714d40d4:  33% ▕█████             ▏  30 MB/ 91 MB   63 KB/s   16m2s
pulling eb2c714d40d4:  46% ▕█████████         ▏  46 MB/ 91 MB   91 KB/s   8m16s

Throughput oscillates between 63 and 154 KB/s, so 91 MB takes fifteen to twenty minutes. A 4 GB 7B model at that rate is hours.

So the real problem in this region is not installing the software, it is getting model files in.

Approach	Note
use smaller models	135m / 0.5B class, tens to hundreds of MB, tolerable
pull elsewhere, copy over	after ollama pull, copy the model directory to the target host
domestic model platforms	ModelScope hosts official model repos, but you must convert to Ollama format — not a one-liner

Everyday commands

ollama list                  # models present locally
ollama pull qwen2.5:0.5b     # fetch a model
ollama run qwen2.5:0.5b      # interactive chat
ollama ps                    # models currently loaded in memory
ollama rm smollm:135m        # delete a model

ollama run opens a REPL; leave with /bye. For one-shot use:

ollama run qwen2.5:0.5b "explain SELinux in one sentence"

Measured CPU inference

Once a model is local, --verbose prints the real per-stage timings:

ollama run smollm:135m --verbose "Say hello in five words."

load duration:        31.149937ms
prompt eval count:    15 token(s)
prompt eval duration: 12.673781ms
prompt eval rate:     1183.55 tokens/s
eval count:           586 token(s)
eval duration:        3.921963228s
eval rate:            149.41 tokens/s

Stage	Meaning	Measured
load duration	model read from disk into memory	31.15 ms
prompt eval	processing the prompt	1183.55 tokens/s
eval	generating tokens	149.41 tokens/s

Those are numbers for a 135M model on CPU alone. Generation speed drops roughly with parameter count, so small models are usable on CPU while a 7B typically manages single-digit tokens per second.

Memory and concurrency

On CPU inference the model is loaded into memory whole. With 3 GB of RAM a 0.5B model is fine; a 7B (about 4.7 GB) will OOM.

free -g | head -2
ollama ps

ollama ps shows the memory each loaded model occupies. When it does not fit, the OOM killer takes the process — visible in journalctl -u ollama -n 50.

Common traps

Symptom	Cause
ollama command missing	only the meta package installed? check rpm -q ollama-base
disk shrank by ~1 GB	ollama-rocm came along; without an AMD GPU install ollama-base
cannot reach 11434 remotely	default bind is 127.0.0.1; set OLLAMA_HOST
model pulls crawl	files come from ollama.com with no regional acceleration
machine freezes on a large model	not enough RAM; the model loads whole
ollama --version prints only a warning	the service is down; systemctl start ollama

Related pages

• Podman in practice — running Ollama in a container instead
• Multi-version installs — several runtimes side by side
• Domestic mirrors — package repos and mirrors
• Official: Ollama documentation