Closed, open weight, open source: what the AI model labels mean, and why I'm changing how my agent runs

On 22 September I switched off my own AI agent.

I run 15 agents and sub-agents on a Mac mini. I turned every job off a few weeks ago after I audited it and found two problems. It had too much access. And nearly everything it did ran on frontier models when they didn’t need to.

The first is a data security story and not all that interesting. This one is about the second, and it comes down to three labels you’ll see on AI models: closed, open weight and open source.

The three labels

Every AI chatbot runs on a model. It’s one very large file of numbers, produced by training on a huge amount of text. Everything else, the app and the chat window and the memory, is packaging around that file. The labels describe who gets to hold the file.

Closed. You never get the file. You rent access through an app or an API (a way for software to send it questions and get answers back). Claude, ChatGPT and Gemini work like this. You get the best quality going and the company handles everything. If they change the price or the model, you change with them.

Open weight. The company publishes the file itself, and you can run it on your own computer. The “weights” are those numbers. Meta’s Llama, Alibaba’s Qwen, DeepSeek, Mistral, Google’s Gemma and OpenAI’s gpt-oss are all open weight. Each comes with its own licence, and you usually don’t get the training data or the code used to build it.

Open source. The strict version. The weights, the training code and enough detail about the data that someone could rebuild it, all under a licence that lets you do almost anything. There are far fewer of these, and plenty of models marketed as “open source” are really open weight.

If it helps, think of a restaurant, a ready meal and a ready meal with the recipe. Closed is the restaurant. Open weight is the ready meal you take home. Open source comes with the recipe too.

Why I started caring

When I looked at what Tommy’s sixteen jobs actually did all day, very little needed the smartest model on Earth. A morning briefing. Sorting notifications. Tidying research notes. Formatting things. I’d already written about paying frontier prices for formatting, and this is the next step: owning the model that does the boring work instead of renting a cheaper one.

The open models have also got a lot better. The best ones now handle everyday jobs like summarising, sorting, drafting and pulling facts out of a page well enough that most people couldn’t tell which one wrote the paragraph. Smaller versions run on a laptop or a Mac mini. A free tool called Ollama makes it one command: type the model’s name, it downloads, and you can talk to it with no account and no bill.

What running one at home gets you

• Cost. A model on hardware you already own costs electricity. For a job that runs every hour, forever, that’s a big difference to a bill that grows with every task.
• Control. The version you downloaded today behaves the same next year. No surprise price rise, no quiet update that breaks your prompts.
• Privacy. Your notes, inbox and files don’t leave the building.
• Resilience. If a provider has an outage, or you hit a usage limit at 11pm, the local model keeps going.

What it costs you

• It’s weaker at the hard stuff. For long careful reasoning, tricky writing, or anything where one wrong step ruins the result, the top closed models are still ahead. I’m not moving those jobs.
• Your machine sets the ceiling. The biggest open models need serious hardware. On a Mac mini I run the smaller ones, and the smaller they are, the more they stumble.
• You become the IT department. Updates, storage, picking the right version, noticing when it’s broken.
• Licences have fine print. “Free to download” doesn’t always mean free to use for anything. Read the licence before you build a business on one.

And the lesson from my audit applies either way. A local model with access to your files and email is just as risky as a rented one if it can act with nobody checking. Owning the model doesn’t make it safe. Permissions do.

The plan for Tommy

Tommy stays off until the clean-up is done. Then it comes back as a mixed team:

• Small open models, running locally through Ollama, for the repetitive jobs: sorting, summarising, formatting, first drafts.
• Bigger open models in the cloud for specific jobs only, each one assigned to named tasks, with guardrails.
• A top closed model for anything that needs real judgement. Anything that sends, posts or pays still waits for me.

None of that is live yet. I’ll write up what happens once it is, including the parts that don’t work.

What I’d tell you

If you only use AI through a chat app, none of this needs to change what you do. The closed models are still the best place to start.

But if you pay every month for AI to do the same dull job over and over, it’s worth knowing there’s a free alternative you can try. Install Ollama, download a small model, and give it one boring task you’d normally hand to ChatGPT. Twenty minutes will tell you if it’s good enough.