Stop 11 · Large Language ModelsTopics · 6/9

Open-Source vs Closed-Source Models

In short

Rent a model through an API, or download one and run it yourself? This choice affects your cost, your privacy story, how much you can customise, and who you call when it breaks.

Closed models (GPT, Claude) are accessed via API only; open-weight models (Llama, Mistral) can be downloaded, fine-tuned and self-hosted — a decision with major cost, control and privacy implications.

The blob stands between a waiting taxi and its own small car with the bonnet open and a toolbox nearby.

Think of it like this

Cab vs your own car

Booking a cab is easy: no servicing, no parking, pay per ride. Owning a car means upfront cost and maintenance, but you go anywhere and can modify it. Closed models are the cab. Open-weight models are the car in your garage.

Watch it happen

Closed API vs open weights

Closed (API)

No servers to manage
Often the strongest models
Data goes to the provider

Open weights

Run on your own hardware
Fine-tune and inspect freely
You handle scaling and updates

Start with an API to prove the idea quickly. Move to open weights when privacy, cost at scale or deep customisation really demands it.

The big ideas

What you need to know

Open weights is not always open source

Many 'open' models share weights but not training data or code, and licences may limit some uses. Read the licence.

Privacy and control

Self-hosting keeps data in your network, which matters for hospitals, banks and government projects.

Hardware is the real cost

Bigger models need serious GPU memory. Quantised small models can run on a laptop, but slower and less capable.

Customisation

With weights you can fine-tune, quantise and change decoding freely. APIs offer only what the provider exposes.

See it in code

Walk through the code

bash

1# Install Ollama from ollama.com, then:2ollama pull llama3.2          # downloads a small open-weight model3ollama run llama3.2 "Explain UPI in two lines."4 5# It also serves a local HTTP API on port 114346curl http://localhost:11434/api/generate \7  -d '{"model": "llama3.2", "prompt": "Hello", "stream": false}'

1/2Download a small Llama model once, then chat with it from the terminal.

Ollama runs open-weight models on your own machine. Nothing leaves your laptop.

Save yourself the pain

Common mistakes

Oops

Assuming self-hosting is always cheaper.

Do this instead

GPUs, engineers and uptime cost money. At low traffic, a pay-per-token API is often cheaper overall.

Oops

Calling every downloadable model 'open source' and ignoring the licence.

Do this instead

Check what the licence allows, especially for commercial use, before building a product on it.

Your mission

Run a model offline

~30 min

Quick check

A hospital wants patient notes summarised but cannot send data outside its network. What's the best fit?

Go deeper

1List the trade-offs: cost, control, privacy, latency, quality for each approach.
2Try running a small open-weight model locally (e.g. via Ollama).
3Learn when self-hosting actually makes financial sense vs. when it doesn't.