Series · Open source AI, explained · Episode 1 / 11

October 11, 2026·8 min read

What is an AI model? Open source vs closed, finally explained.

Within a year, people who know how to run AI on their own machine will have a head start. Most people don’t even know it’s possible. This guide accompanies episode 1 of the series: keep it handy, you’ll need it for what comes next.

Idriss Laouali

Co-founder, Karatou Academy

AI · Open source · Beginner · Series

1. AI has two parts

The model = the brain

A file (often a few gigabytes) containing what the AI learned from vast amounts of text. Examples: GPT, Claude, Llama, Qwen. Source

The app = the window

The interface where you type, with buttons and your conversation history. Example: the ChatGPT app. Source

ChatGPT is the app. GPT is the model.

2. Closed AI or open AI?

Closed AI

The brain stays on the company’s servers. You send your question over the internet, and never get the model file.

Your questionInternet → their servers

ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google). Source

Open AI

The company publishes the model file. You download it and run it at home, even without the internet.

The file → your machineThen offline

Llama, Gemma, Qwen, Mistral, DeepSeek, gpt-oss. Source

Closed and open AI compared
Criterion	Closed	Open
Where does the brain run?	On their servers	On your machine
Internet needed?	Yes	No, once downloaded
Cost	Limited free access or a subscription	Free (you need suitable hardware)
Your data	Sent to the company	Stay on your machine
Customizable?	Limited	Yes, you can adapt it (languages, work)

Here we compare online services with a model running entirely locally. An open model can also be hosted online: your data stays on your machine only if the app does not send it to an external service. Source

3. The main open models

Llama

Meta · United States

Llama Community License (specific conditions) Source

Gemma

Google · United States

Gemma Terms of Use Source

Qwen

Alibaba · China

Apache 2.0 for most versions Source

Mistral

Mistral AI · France

Apache 2.0 for several models Source

DeepSeek

DeepSeek · China

MIT license (DeepSeek-R1) Source

gpt-oss

OpenAI · United States

Apache 2.0 (August 2025) Source

Main license: varies by version. Always check the license of the file you download.

Even OpenAI, the company behind ChatGPT, has been releasing open models since August 2025: gpt-oss. Source

4. The nuance: open weights does not always mean open source

Most of these models have open weights: you get the brain (the file), but not always the recipe (training data and training code). In 2024, the Open Source Initiative published a definition of open source AI requiring more transparency. For you, the key point is that the brain is on your machine, and you can use it within the terms of its license. Source

The brain: provided

The recipe (training data): not always

5. Good habits before downloading

Download only from official sources (the company’s website, Hugging Face, the Ollama library, LM Studio), never from a link forwarded on WhatsApp.Source (Hugging Face)Source (Ollama)Source (LM Studio)
Read the license if you plan to use it in a product or business.Source
Check the model size and your machine’s memory (RAM). We cover this in episode 3.Source
Open models hallucinate too: always check important answers.Source

A short glossary

Model: The file containing what the AI has learned and used to generate its answers.
Parameters: Values learned during training. “B” means billion: 8B = 8 billion parameters.
Open weights: The model’s learned parameters are available to download under its license terms.
License: The rules defining what you can do with a model.
Offline (local): The model runs on your machine without sending every question to a server.

Source (Alibaba)Source (Open Source Initiative)Source (LM Studio)

The series

1
What is a model? Open source vs closed
Available
2
Why it's a huge opportunity for AfricaAvailable
3
8B, 70B: choosing the right model
Coming soon
4
The limits, honestly
Coming soon
5
Install AI on your computer with LM Studio
Coming soon
6
Ollama, the developer option
Coming soon
7
Which model for your machine?
Coming soon
8
Android: Google AI Edge Gallery
Coming soon
9
iPhone and Android: PocketPal AI
Coming soon
10
The airplane-mode test
Coming soon
11
AI in our own languages
Coming soon

Join the community

The WhatsApp channel is where I share the guide, resources and every new episode. Coming soon: a platform where you can contribute your voice and help build AI for Niger in our languages.

Join the WhatsApp channel

Questions? Leave them in the comments under the video. I answer every one.

Sources

OpenAI, ChatGPThttps://openai.com/chatgpt/overview/
Meta, Llama Community Licensehttps://github.com/meta-llama/llama-models/blob/main/models/llama3_3/LICENSE
Google, Gemma Terms of Usehttps://ai.google.dev/gemma/terms
Alibaba, Qwen3 models and licenseshttps://github.com/QwenLM/Qwen3
Mistral AI, Mistral 7B (Apache 2.0)https://mistral.ai/news/announcing-mistral-7b/
DeepSeek, DeepSeek-R1 and licensehttps://github.com/deepseek-ai/DeepSeek-R1
OpenAI, Introducing gpt-oss (August 2025)https://openai.com/index/introducing-gpt-oss/
Lelapa AI, InkubaLM (2024)https://arxiv.org/html/2408.17024v1
Nigeria, N-ATLAS at UNGA80 (2025)https://www.opensourceforu.com/2025/09/nigeria-unveils-open-source-n-atlas-ai-model-to-champion-african-languages-at-unga80/
Open Source Initiative, Open Source AI Definition 1.0https://opensource.org/ai/open-source-ai-definition
Ollama, official model libraryhttps://ollama.com/library
LM Studio, offline operationhttps://lmstudio.ai/docs/app/offline
Hugging Face, model hub and model cardshttps://huggingface.co/docs/hub/model-cards
OpenAI, Why language models hallucinatehttps://openai.com/index/why-language-models-hallucinate/

Idriss Laouali

Co-founder, Karatou Academy