IntegrAuth Academy
Identity, security and practical AI — free, from zero. Byte-sized lessons, real diagrams, hands-on labs, zero fluff. Two programs, each with its own exam and certificate.
English is hard? Your browser can translate this page into Tamil, Telugu, Kannada, Malayalam or Hindi. Tech words stay in English so you learn the real names.
How?
- Chrome on phone: ⋮ menu → Translate
- Chrome on computer: right-click → Translate to …
- Edge: the translate icon in the address bar
Open and closed models
Priya's legal team wants AI help with contracts. The contracts must stay in the building. Your goal: learn what "open" means, and what fits on a laptop.
🧪 Interactive lab — enable JavaScript to play with this one.
- A closed model stays on its maker's servers. You can download an open-weight model.
- Free to use is not the same as open source.
- Memory needed ≈ parameters × bits ÷ 8, plus extra room.
What "open" means
Apps are made from written instructions, called source code. Some code is free to use, study, change and share, even for business. This is open source. A legal text, the license, gives these rights.
A model's main part is its weights: billions of numbers learned in training. A closed model keeps them on the maker's servers. An open-weight model lets you download the weights and run them yourself.
Will it fit?
The weights must fit in your computer's memory: parameters × bits ÷ 8 = GB. An 8-billion-parameter model at 16 bits needs 8 × 16 ÷ 8 = 16 GB. The chat and the system need extra room.
You can store each number with fewer bits. This is quantization. At 4 bits, the same model needs about 4 GB. Quality usually drops a little.
Run it yourself?
✅ Good reasons
Private text stays with you. It works offline. Heavy daily use has a steady cost. No one can retire your model.
⚠️ Think twice
The best models are usually closed or too big. You handle updates, security and growth. For light use, paying a provider per use through an API is often cheaper.
Now check some downloads.
🧪 Interactive lab — enable JavaScript to play with this one.
Go deeper: licenses
The Open Source Initiative keeps the common definition of open source, and a list of approved licenses.
| License | In plain words | Type |
|---|---|---|
| MIT | Do almost anything, even sell it. Keep the copyright notice. | Permissive |
| Apache-2.0 | Like MIT. It also gives patent rights. You must note your changes. | Permissive |
| GPL | Share a changed version? Then share your changes under the same license. | Copyleft |
| Custom model license | Written by one company. Often free, but with conditions, like size limits, banned uses or a required credit line. | Not open source by the usual definition |
As of October 2026, gpt-oss, most Qwen3 models and Mistral 7B use Apache-2.0. DeepSeek-R1 uses MIT. Meta's Llama models have their own license with conditions. A model family can change its license between versions.
A model card is the fact sheet that comes with a model, for example on Hugging Face. It names the license. Read it before you build something to sell.
Go deeper: fully open AI
An AI model has four parts: the code to run it, the weights, the training data and the training code. So ask: which parts can I get, and under what license?
Fully open-source AI shares all you need to study and rebuild the model. The Open Source Initiative made a definition for this in October 2024. It asks for the weights, the full code and details about the training data, all under open terms. Few models meet it. Most open-weight models do not share their training data. Ai2's OLMo family is a well-known example that does. It shares its data, code, weights and training checkpoints.
Go deeper: memory math
The weights sit in memory while the model runs. This can be the computer's main memory, called RAM. It can also be the graphics card's own memory, called VRAM. VRAM is much faster for this work.
A 70-billion-parameter model at 16 bits needs 140 GB.
8-bit is usually very close to full quality. 4-bit is a popular middle choice for laptops. Below 4 bits, quality often falls more, depending on the model. File names often show the level, like Q8_0 or Q4_K_M in the common GGUF format.
Go deeper: Priya's setup
Free tools can run a model on your machine with no internet. Ollama and llama.cpp are open source. LM Studio is a free desktop app, but it is not open source. These tools download a model file, then let you chat with it.
Priya's team picked an 8-billion-parameter open-weight model with a permissive license. They run it at 4 bits in Ollama, on laptops with 32 GB of memory. It does a good first pass. A small local model is weaker on hard tasks, so the lawyers still check every answer. For hard questions on files that are not secret, they use a bigger closed model.
With an open model, you can also change the model itself. See Prompt, RAG or fine-tune? A local model with tools still needs limits. It can still fall for prompt injection.
A model file is software from the internet, so it is a supply-chain risk. Some old formats, like Python "pickle" files, can run hidden code when they load. Use safetensors or GGUF files from publishers you trust.
Install Ollama. Get a small model at 4 bits. Turn off Wi-Fi and chat with it. Then find its license on its model card.