INQUIRE Lab · Educational

The Open LLM Explorer

A friendly, no-jargon tour of the open-source Large Language Models you can actually download and run yourself — who builds them, how big they are, and how to get one running on your own laptop.

New here? Start with the basics

You don't need any AI background. A few short ideas and you'll understand everything below.

💬

What is an LLM?

A Large Language Model is a program trained on huge amounts of text. Its one job: read what came before and predict the next word (really the next token — a word-piece). Repeat that and it can write, answer, summarize and code.

Text in, text out. That's the whole idea.

🔢

"Parameters": million vs billion

Parameters are the model's learned settings — millions of tiny knobs tuned during training. More knobs = more capacity for patterns.

1M = a million. 1B = a billion (1,000 M). A 7B model has 7 billion knobs; a 70B has ten times more.

Rule of thumb: bigger is usually smarter, but needs more memory and runs slower. Bigger isn't always better for a task.

🧱

The architecture, in 60 seconds

Almost all of these are Transformers. Text → tokens → vectors → stacked layers of attention (each token "looks at" the others) → a prediction for the next token.

Many 2026 models are Mixture-of-Experts (MoE): huge in total parameters, but only a small "active" slice runs per token — big brain, fast answers.

🔓

Open-weight vs closed

Open-weight: download it and run offline, free (Llama, Gemma, Qwen, DeepSeek…). Fully open adds the data + code (OLMo).

Closed (GPT, Claude, Gemini) live only behind an API. This page is about the open ones you can own.

How big is "big"? A sense of scale

Roughly how much memory each size needs locally (using common 4-bit "quantization", which shrinks models to fit consumer hardware).

~1Bphone
3–4B~3 GB
7–9B~5 GB
27–32B~18 GB
70B~40 GB
235B+ MoEserver

A 16 GB laptop runs 7–9B models happily. A 24 GB GPU handles ~30B. The giant MoE models want a server — but their smaller siblings run anywhere.

The models, by company

Each card is a company / model family with its open models. Tap any model for details and a one-line command. Every icon is downloadable.

chat coding reasoning tiny (runs anywhere) multimodal
license size good for

Run one on your own computer

Two beginner-friendly tools. Both download a model once, then run it fully offline and private.

Ollama (terminal, super simple)

Install from ollama.com, then in a terminal:

ollama run llama3.3
ollama run gemma3
ollama run qwen3
ollama run deepseek-r1 # reasoning
ollama run phi4 # tiny + smart

Other handy commands:

ollama pull qwen3:7b # download only
ollama list # what you have

An OpenAI-compatible API is served at localhost:11434.

LM Studio (friendly app + CLI)

A desktop app from lmstudio.ai: search a model, click Download, then Chat — no terminal needed.

  1. Open LM Studio → 🔍 search (e.g. "Llama 3.1 8B").
  2. Pick a quant (Q4 is a great default) → Download.
  3. Go to Chat, load the model, start typing.

Prefer the terminal? LM Studio ships the lms CLI:

lms get llama-3.1-8b # download
lms load llama-3.1-8b # load it
lms chat # chat in terminal
lms server start # local API

Its local server runs at localhost:1234.