AI Engineer Guide

The AI Model
Landscape.

A tour of the pretrained models you will actually use: text, image, audio, and video. Who makes them, which are open or closed, and what each one is good for.

28 models

4 modalities

2 access types

01 · The two big distinctions

Open or closed. API or self‑hosted.

Almost everything you build runs on a model someone else already trained, at a cost of millions of dollars. You do not build these, you pick one. Two distinctions explain the whole map.

Closed

owned by a company

Reached only through a paid API. You cannot see or change the internals. You rent the intelligence.

GPT Claude Gemini

+Top quality, zero setup, scales instantly
–Pay per use, no control, data leaves your machine

Open

weights you can download

The weights are public. You can run them, inspect them, and fine-tune them yourself. You own the intelligence.

Llama Qwen DeepSeek

+Full control, private, no per-call cost
–You need the hardware and the setup know-how

API

Someone else runs the model, you call it over the internet. Fastest way to start. You pay per use. Every closed model and most open ones are available this way.

Self‑hosted

You run an open model on your own machine. Full control, no per-call cost, but you need the hardware. The easiest way in is Ollama.

02 · Pick your tool

The model explorer

Filter by what you want to build and how you want to access it, then hover any model for its card or click through to the docs. You don't need to learn them all: start with the text models, where most AI engineering happens, and treat the rest as a reference for when a project needs image, audio, or video.

Hover a model to open its card · click to visit the docs

Where to find them

Hugging Face

Think of it as GitHub but for AI models. It hosts hundreds of thousands of open models, datasets and tools, and it is where you will read the model card and check the real file size before you download anything. Every AI engineer ends up here eventually, so the sooner you get familiar, the better.

huggingface.co/models →

Where to compare them

Live leaderboards

Rankings shift every few weeks, so do not memorize a single “best” model. Learn where to look instead: Artificial Analysis scores quality, price and speed side by side, and LMArena ranks models by blind human votes.

Artificial Analysis → LMArena →

03 · Running one yourself

What the numbers on an open model mean

With a closed model you just call the API. The moment you download an open one, a row of numbers shows up in its name and you have to work out what your machine can actually handle. Five ideas cover almost all of it.

Model size

The B is billions of parameters, the numbers the model learned during training. More parameters usually means smarter, and always means heavier: every one of them has to sit in memory while the model runs. This is the first number to look at, because it decides whether the model fits on your machine at all.

Qwen3-8B → 8 billion parameters

Active parameters (MoE)

A Mixture of Experts model is split into many small expert networks and wakes only a few of them per token. So it has two sizes, and they do different jobs: total parameters decide the memory, active parameters decide the speed. You need room for all of it, but you only pay the compute of the part that runs.

Qwen3-30B-A3B → 30B in memory, 3B per token

Quantization

Weights are trained at 16 bits each, but they can be stored at 8, 4, or fewer. That is quantization, and it is what makes local models practical: 4-bit roughly quarters the file for a small quality loss. Q8 is nearly lossless, Q4 is the usual sweet spot, and below 4-bit quality drops off quickly. Those names come from GGUF, the file format Ollama and llama.cpp read.

Q4_K_M → an 8B model is ≈5 GB, not 16 GB

Context size

You already know the context window is the model’s working memory (see Core LLM elements). What is easy to miss is that holding it costs memory of its own, on top of the weights, and that cost grows with every token in the conversation. This is why local runtimes ship with a small context by default and make you raise it deliberately.

longer conversation → more memory, same model

Will it run, and will it be fast?

Two questions, and both come down to memory, not to how powerful your processor is.

Will it run? The model and the conversation both have to fit in your memory at the same time. If they do not fit, it simply will not start. So count the size of the model, then leave room on top for the chat.

Will it be fast? To write each single word, the model has to read all its active weights again. So the speed you get depends on how fast your machine can move that data around. A big model on a slow machine still answers, just one slow word at a time.

Will it run? weights + context ≤ your memory

How fast? how fast your memory is

Rough guide: at Q4, expect a bit over half a gigabyte for every billion parameters, plus some room for the chat. It changes from model to model, so check the real file size on Hugging Face before you download.

Sources & further reading Artificial Analysis, leaderboards by quality, price and speed · artificialanalysis.ai
LMArena, rankings by blind human votes · lmarena.ai
Hugging Face, the hub for open models · huggingface.co/models
Ollama, run open models locally · ollama.com
llama.cpp, the engine behind GGUF quantization · github.com/ggml-org/llama.cpp