Blog / Build with AI /
Open AI models: Llama, Mistral, Qwen and why they matter
Most people meet AI through a website: you type, a company’s servers think, an answer appears. But there’s a second kind of AI model that you can download, keep on your own machine and run without asking anyone’s permission. Llama, Mistral and Qwen are the best-known families of these.
Here’s what that means in plain words, why it matters, and what it changes if you build things or run a website.
What is an “open” model?
Every AI model, in the end, is a very large file of numbers, called the weights. A closed model, like the one behind ChatGPT, lives on the company’s servers and you only reach it through their app or paid API. An open-weights model is one where the company publishes that file so anyone can download it.
Think of the difference between eating at a restaurant and being handed the recipe plus the ingredients. At the restaurant you get the meal, but only on their terms and at their price. With the recipe, you can cook at home, tweak it, and keep cooking if the restaurant closes.
One honest caveat: “open” doesn’t always mean “open source” in the strict sense. Some models come with licences that limit certain uses, such as very large commercial deployments, and the training data is usually not shared. Read the licence before you build a business on one.
The three names you’ll hear
Llama (Meta)
Meta’s Llama family is probably the most widely used. Llama 3.1, released in the summer of 2024, came in several sizes, including a very large 405-billion-parameter version, and the smaller ones run on ordinary hardware. Many other tools and fine-tuned models are built on top of it.
Mistral
Mistral AI is a French company. Its first small model, Mistral 7B, made people sit up in 2023 because it did a lot for its size. It has released a range of models since, some with open weights and others only through its paid service.
Qwen
Qwen comes from Alibaba. The Qwen2.5 series, from late 2024, includes a wide range of sizes and is well regarded for coding and for working across many languages.
I’d also mention DeepSeek, whose R1 reasoning model made headlines a few weeks ago. I covered why in DeepSeek R1: why a cheap AI model shook the industry.
Model names and versions change fast, so treat this list as a snapshot, not a ranking.
Why open models matter
1. Privacy
When a model runs on your own computer, your text doesn’t leave it. If you handle client documents, customer lists or anything sensitive, that is a real advantage. Local AI models: running an LLM on your own computer shows what it takes.
2. Cost
Running a model locally has no per-question fee. You pay for the hardware and electricity instead. For lots of small, repetitive tasks (tagging, summarising, cleaning data) that can be cheaper than paying for an API. For occasional use, a paid API is usually cheaper than buying a powerful computer.
3. Control and independence
A hosted model can change overnight: its behaviour, its price, its rules, or whether you can use it in your country. An open model on your machine stays exactly as it was when you downloaded it. That stability matters if you build a workflow around it.
4. Customisation
Because you have the weights, you can fine-tune an open model on your own data, or run a version specialised for one job. That’s out of reach with most closed models.
5. Competition
Open models put pressure on prices and keep the big companies honest. When a good model is free to download, the paid ones have to justify their price.
What open models don’t do
- They’re not automatically better. The largest closed models are still ahead on many hard tasks. The open ones are often “good enough” for everyday work, and that’s a different claim.
- They’re not free to run at scale. A big model needs serious hardware. Small ones run on a decent laptop, but they’re less capable.
- They don’t remove the need to check. Open models make mistakes and invent facts like any other. See evaluating AI output: simple tests for quality.
- They’re not immune from licence problems. Check the terms.
How to try one without a headache
You don’t need to be a developer. Two tools make it simple:
- Ollama runs models from a simple command line or app.
- LM Studio gives you a point-and-click window to download and chat with models.
Start with a small model (a few billion parameters) and see how it feels. If your computer struggles, that’s the answer to “do I need a bigger machine?” For choosing hardware, laptops for running local AI models covers what matters, mostly memory.
What it means if you run websites
A few practical uses:
- Drafting product descriptions or meta titles in bulk, without sending your catalogue to a third party.
- Classifying support emails or comments.
- Summarising long documents or logs.
- Building a simple chatbot on your own documents. See RAG explained: giving AI your own documents.
For each one, the question isn’t “is an open model cool?” It’s “does this job need a top-tier model, or is a smaller one fine?” Often a smaller one is fine. That’s the idea behind cheaper models for simpler tasks.
Key takeaways
- Open-weights models are files you can download and run yourself.
- Llama (Meta), Mistral and Qwen (Alibaba) are the main families to know.
- The big wins are privacy, control, predictable cost and the ability to customise.
- They aren’t automatically smarter, and licences differ, so read them.
- Try a small one with Ollama or LM Studio before you decide anything.
Related reading
Would you use a model that lives on your own laptop, or is a hosted one easier for you? Tell me which jobs you’d give it.