# Free, Open and on Your Laptop: Open-Source Models and the Secret Currency Called Tokens

> Part 5 of AI From Zero to Agents. What open-source (really, open-weight) AI models are, how to run a capable one on your own laptop with no internet, and tokens, the unit behind every AI limit and bill. Plus why the same sentence costs more tokens in Arabic than in English.

- Author: Maged Zamzam (https://magedzamzam.ae/profile)
- Published: 2026-11-04
- Series: AI From Zero to Agents, part 5
- Topics: AI, Open source AI, Tokens, LLM, Local AI, AI for beginners
- Canonical: https://magedzamzam.ae/notes/open-source-models-and-tokens-explained

---
There are two ideas sitting underneath nearly every AI decision a company makes, and I rarely see anyone explain them to non-technical people. The first is whether a model is open or closed, in other words whether you can download it and run it yourself or can only rent it. The second is tokens, the unit that every AI limit, speed and bill is measured in.

Once these two click, you can hold your own in pretty much any conversation about AI cost, privacy or strategy. It's also where [Part 4](https://magedzamzam.ae/notes/chatgpt-vs-claude-vs-gemini-vs-grok-which-ai-to-use) left off, when I said Meta's Llama is "open". So what does that actually mean?

## Open models

### Closed vs open

![Comparison: a closed model is like a restaurant, you order through an app or API, the vendor keeps the recipe and the kitchen, and your data travels to them; an open-weight model is like a recipe book, you download the model, run it on your own laptop, server or cloud, and your data stays with you](https://magedzamzam.ae/media/bbbb76afd490a9d9ebfd.jpg "Closed models are a restaurant. Open models are a recipe you cook at home.")

The way I usually explain it is restaurants versus recipes. Closed models such as GPT, Claude, Gemini and Grok only run on the vendor's own computers. You order through their app, or pay per use through their API, and the kitchen stays theirs. You get top quality and no setup at all, but your data travels to them, and you live with their prices and their decisions.

Open-weight models hand you the recipe. The vendor publishes the trained file itself (the "weights", or parameters, from Part 3), and you can run it wherever you like: your laptop, your company's servers, any cloud you fancy. You get control, privacy and no per-question fee. In exchange, you provide the hardware and look after it.

### "Open source" isn't quite the right word

Strictly speaking, most "open-source" AI is open-weight. You get the trained model, but not the training data or the full method behind it. It's a bit like being given a cake and its ingredients list without access to the bakery. Licences vary too: some are completely permissive (Apache 2.0, MIT), others add conditions, such as limits for very large companies. So please read the licence before you build a product on one. It's boring, I know.

### Who makes them

![A map of major open-weight model families grouped by origin: from the US, Meta's Llama, Google's Gemma and OpenAI's gpt-oss; from Europe, Mistral; from China, DeepSeek, Qwen, Kimi and GLM](https://magedzamzam.ae/media/4bfb04ae2de146007f32.jpg "The main open-weight families, October 2026")

Two things tend to surprise people here. One is how close the open models are. The best of them usually trail the closed frontier by months rather than years, and for a lot of everyday tasks you genuinely can't tell the difference. The other is that even the closed labs publish some: OpenAI's gpt-oss (under an Apache 2.0 licence) and Google's Gemma sit alongside their paid flagships.

Most of this lives on Hugging Face, which people like to call the GitHub of AI. It hosts well over a million models you can download.

![The Hugging Face models page listing millions of downloadable models with filters by task, library and licence](https://magedzamzam.ae/media/0683c88604ec64dfef19.jpg "Hugging Face, the public library of open AI models. Screenshot: huggingface.co, October 2026.")

### Yes, it runs on your laptop

This is usually the point in a workshop where people sit up. With a free app like Ollama or LM Studio, a recent laptop with 16 GB of memory can run a decent small model completely offline. No account, no internet, and nothing leaves your machine.

![Three steps to local AI: install a free app such as Ollama or LM Studio, choose a small open model such as a Gemma, Llama, Qwen or gpt-oss variant sized for your memory, then chat offline with your data never leaving the laptop](https://magedzamzam.ae/media/547d2fd6df2c32fcf92d.jpg "Private AI in three steps")

![The Ollama model library website showing popular open models available to download and run locally](https://magedzamzam.ae/media/0aeadcc69b9e5b182cfa.jpg "Ollama’s model library: pick a model, run one command, chat offline. Screenshot: ollama.com, October 2026.")

How big a model can you run? Roughly this:

| Your machine | Comfortable model size | What it feels like |
|---|---|---|
| Phone / 8 GB laptop | 1–4 billion parameters | Quick helper, simple drafting and summaries |
| 16 GB laptop | 7–20 billion | Solid assistant for everyday writing and Q&A |
| 32–64 GB workstation or Mac | 30–70 billion | Close to last year's flagships |
| GPU servers | 100+ billion | Frontier-class open models, for companies |

What makes this possible is a trick called quantisation. The model's numbers are stored with less precision, much like saving a photo as a smaller JPEG. You lose a little quality and keep most of the ability.

So when does running locally make sense? When privacy matters (patient notes, legal drafts, customer data), when you're offline, or when you're going to run millions of small jobs and per-use fees would pile up. For the really hard problems, the closed flagships are still ahead.

## Tokens, the secret currency

### What a token is

Models don't read words. They read tokens, which are chunks of text: often a whole short word, sometimes a piece of a longer one, a space plus a word, or a lone full stop. Rather than describe it, I ran a few real sentences through the tokenizer that OpenAI's recent models use. This is what the model actually sees:

![The English sentence “Artificial intelligence is changing how we work in Dubai.” split into 10 tokens, and the same sentence in Arabic split into 16 tokens, plus the word “Unbelievably” split into three tokens: Un, bel, ievably](https://magedzamzam.ae/media/0fc350feaf05f3fde8ca.jpg "Real tokenizer output: the same sentence costs 10 tokens in English and 16 in Arabic")

For English, a decent rule of thumb is that one token is about three quarters of a word, or roughly four characters. So 1,000 tokens comes to around 750 words.

### Why Arabic costs more

Look at the Arabic line again. Seven words turned into 16 tokens, while nine English words became just 10. Tokenizers are mostly built from English-heavy data, so common English words get a token of their own while Arabic words often get chopped into pieces. For the same meaning, Arabic text uses more tokens, which means it costs more and fills up the context window faster. The gap varies from model to model, and newer tokenizers have narrowed it a lot, but it's still there.

If you're building Arabic AI products here in the Gulf, measure this on your own content. It can change the business case.

### Why you should care

![Four everyday effects of tokens: limits, as free plans cap messages and context by tokens; cost, as APIs charge per million tokens with output costing more than input; speed, as answers stream at a number of tokens per second; and memory, as the context window from Part 2 is measured in tokens](https://magedzamzam.ae/media/2ca8a3cab17c3e27e0e4.jpg "Four places you meet tokens without noticing")

You bump into tokens all the time without realising. Free plans and context windows are counted in them. Answers stream onto your screen at so many tokens per second. The context window from Part 2 is really a token budget. And businesses using AI through an API pay per million tokens, with one twist worth remembering: output tokens cost several times more than input tokens, because writing is more work than reading.

### What a million tokens buys

![An infographic of what one million tokens is roughly equal to: about 750,000 English words, about ten average novels, about 2,500 pages of a report, or about 1,000 typical chat conversations](https://magedzamzam.ae/media/84d796c3be8b9e8dee87.jpg "One million tokens, in human terms (English, approximate)")

Here's a worked example. Summarising a 10-page report (about 5,000 words in, 300 words out) takes roughly 6,700 input tokens and 400 output tokens. At US$2 per million input tokens and US$6 per million output, which is Grok 4.7's published list price in October 2026 and a fair mid-range figure, that comes to under one US cent. Prices swing a lot (small models cost a fraction of that, top flagships several times more), but it explains why companies can afford to point AI at millions of documents.

## A quick glossary

Since this part threw a lot of new words at you:

- Open-weight model: a model whose trained parameters you can download and run yourself.
- Licence: the legal terms that say what you're allowed to do with an open model.
- Quantisation: compressing a model so it fits on smaller hardware.
- Token: the chunk of text a model reads and writes, and the unit AI is priced in.
- API: how software (rather than a person in a chat window) uses an AI model, usually paid per token.

## Two things to try

First, go to platform.openai.com/tokenizer, or any "token counter" site, and paste in a paragraph in English, then the same paragraph in Arabic or your own language. Compare the counts and you'll see the effect for yourself.

Second, if you're feeling brave, install Ollama or LM Studio, download a small model, turn off your Wi-Fi and ask it something. That's private AI, running entirely on your own machine, and it still feels slightly magical to me.

In Part 6 we hand the model some tools, a goal and a loop, and it stops just talking and starts actually doing things.

---

Source: https://magedzamzam.ae/notes/open-source-models-and-tokens-explained. Views are the author's own.
