Ollama vs LM Studio: which one for your old PC

Line-by-line comparison

No absolute winner: it changes based on how much memory you have and what you need. The "on old PC" column is the only one that matters for those here.

Which of the two for your PC

Select your hardware and the answer takes into account the memory you actually have, not the generic rule.

Step-by-step installation

Basic steps and links to the official documentation. Requirements and installation options can change by system and version.

LM Studio

1. Download the version for your system from lmstudio.ai/download and check the published requirements.
2. Install with the default settings. On first launch choose the "My Documents" mode only if the system disk is almost full: the models are large.
3. Look for a model suited to your RAM and your use; the download size and the time depend on the variant and the connection.
4. In the menu on the left open "Local Server" and press Start. LM Studio exposes a server compatible with the OpenAI APIs on localhost:1234: any program that accepts an OpenAI endpoint can use it.
5. Start the model with a context suited to the available memory and check the RAM usage during a test.

Ollama

1. Windows or macOS: download OllamaSetup.exe from ollama.com/download and install it. On Windows it becomes a background application with an icon in the status bar.
2. Linux: a single command, which installs the binary and creates the system service:

   curl -fsSL https://ollama.com/install.sh | sh
   sudo systemctl enable --now ollama

3. Download and start a model with a command:

   ollama run qwen2.5:7b

4. For everyday use it is better to preload the model without starting it: ollama pull qwen2.5:7b. That way the next startup is immediate.
5. List the installed models with ollama list and delete one you do not use with ollama rm nome-modello: every model takes gigabytes that do not come back.

Using Ollama from other programs

This is the main reason people choose Ollama: a local model becomes usable from any program, even one written in any language, without an account or costs.

The minimal command

Ollama listens on port 11434 of the localhost address. The simplest request is made with curl:

curl http://localhost:11434/api/chat \
  -d '{
    "model": "qwen2.5:7b",
    "messages": [{"role": "user", "content": "Riassumi il teorema di Bayes in tre righe"}],
    "stream": false
  }'

With "stream": false you receive a single JSON block when the model has finished. With true, which is the default, you receive one piece of the response at a time, exactly like in a chat: for a web interface that is what you need.

OpenAI-compatible endpoints

Ollama also exposes a /v1 path compatible with some OpenAI APIs. Check the parameters supported by your program and do not expose the port to the Internet: limit access to the trusted local network or add an appropriate authentication layer.

Python example

With the official ollama library the request is three lines:

import ollama

risposta = ollama.chat(
    model="qwen2.5:7b",
    messages=[{"role": "user", "content": "Ciao, chi sei?"}]
)
print(risposta["message"]["content"])

Confirming it works

The quickest test to understand whether the service is running, with a command that works on Windows, Linux and macOS:

curl http://localhost:11434/api/tags

If it replies with a JSON that lists the models you downloaded, everything is in order. If you get a connection error, the service is not running: on Linux check with systemctl status ollama.

One model, two PCs

In a home with more than one computer, the old PC is not there to run models: it is there to act as a server for the others. It is a use that works even on 2012 hardware.

Ollama is designed for this: just expose it on the local network and every PC in the house will be able to ask it for answers. The model runs only once, on a single machine, and the others take advantage of it.

1. Choose the machine that hosts the model. Better a mini PC or a laptop with 16 GB and even a modest GPU: the GPU matters because inference is tied to memory bandwidth.
2. Make it reachable from the other PCs. On macOS and Windows the setting is in the application options. On Linux, by default restricted to localhost, you must expose it explicitly:

   sudo systemctl edit ollama
   
   [Service]
   Environment="OLLAMA_HOST=0.0.0.0:11434"

   then sudo systemctl daemon-reload && sudo systemctl restart ollama.
3. Open the port in the firewall only on the local network, never on the Internet:

   sudo ufw allow from 192.168.1.0/24 to any port 11434

4. Give the other PCs the address of the machine that hosts the model: if its IP is 192.168.1.50, the other programs will use http://192.168.1.50:11434/v1.

Common problems

Four errors you run into almost always, with the real cause and the solution.

Frequently asked questions