The families of models

A family, not a model

Llama, Gemma, Qwen, Mistral, Phi. These are family names, not single products. One name covers a set of models built by one team from one recipe.

A family usually comes as a ladder of sizes. The Qwen3 report describes a range that runs from 0.6 billion parameters up to 235 billion. You pick the rung your machine can hold.

A family also spreads sideways. Google's pages list Gemma 4 for text, audio and images, EmbeddingGemma for turning text into numbers for search, and ShieldGemma 2 for checking whether content is safe.

The same pages list more members still, among them PaliGemma, FunctionGemma and RecurrentGemma. Google says many more variants exist beyond those.

So the family name tells you the workshop. The rest of the label tells you which model came out of it.

Backed by Qwen3 Technical Report, Gemma models overview.

Who builds them

Almost everything you will run at home comes from a short list of makers.

Watch for one thing as you read it. Some makers reach for a licence that ordinary software already uses. Others write their own. That habit is one of the steadiest facts about a maker, and lesson seven is built on it.

• Meta builds Llama. It carries Meta's own licence, with conditions attached.
• Google DeepMind builds Gemma, which Google describes as built from the same research and technology as its Gemini models.
• The Qwen team builds Qwen, a wide ladder from 0.6 billion parameters up to 235 billion.
• Mistral AI, a French company, builds Mistral. Some of its models are Apache-2.0 and some are for research only.
• Microsoft Research builds Phi, aimed at small models made capable through careful training data.
• DeepSeek builds DeepSeek. Its R1 model holds 671 billion parameters, with 37 billion active at a time, under an MIT licence.
• OpenAI publishes gpt-oss at 21 billion and 117 billion parameters, under Apache-2.0.

Backed by Llama 3.3 Community License Agreement, Gemma models overview, Qwen3 Technical Report, Mistral-7B-Instruct-v0.3 model card, Mistral AI Research License, Phi-4 model card, DeepSeek-R1 model card, gpt-oss-20b model card.

Reading the name

A model's name is not decoration. It follows a rough grammar, and reading it tells you a lot before you open the page.

Take Qwen3.5-0.8B. Qwen3.5 is the family and the version. The 0.8B is the size: eight hundred million weights.

That card names a base model as well, Qwen3.5-0.8B-Base. So the plain name is the tuned one, and the Base name is the raw model it came from. Lesson ten covers why that difference matters.

Now look at Gemma 4. Google ships it in five sizes: E2B, E4B, 12B, 26B A4B and 31B. Two of those carry a letter next to the number.

Google's card explains both letters. The E in E2B stands for effective parameters. The A in 26B A4B stands for active parameters, which the next section is about.

No convention is obeyed everywhere. Some names carry a date, some a word like Instruct or Thinking. Read the parts you know and check the card for the rest.

Backed by Qwen3.5-0.8B model card, Gemma 4 model card: sizes, dense and MoE variants, active parameters.

Two size numbers, doing different jobs

Some models put every weight to work on every word. Others switch most of them off and use a small part at a time.

The second kind is called a mixture of experts. It has two sizes, and each one answers a different question.

The total is what has to fit in your memory. The active part is what sets the speed. A model can be quick to run and still far too large to load.

DeepSeek-R1 shows the gap at its widest. Its card gives 671 billion weights in total and 37 billion active. You still need room for all 671 billion.

OpenAI's gpt-oss-20b is gentler. It holds 21 billion weights with 3.6 billion active, and OpenAI says it runs within 16 GB of memory.

Google's 26B A4B is a third case. Its card gives 25.2 billion weights with 3.8 billion active, and says it runs almost as fast as a 4 billion model.

Backed by DeepSeek-R1 model card, gpt-oss-20b model card, Gemma 4 model card: sizes, dense and MoE variants, active parameters.

The four things that actually differ

Set two families side by side and four differences do all the real work.

• The licence. Some are Apache-2.0 or MIT. Some carry the maker's own terms. This decides what you may do, and it is the next lesson.
• The ladder of sizes. A family with a 0.6 billion rung is one you can try on a small machine. One that starts much higher is not.
• What it is built for. General chat, code, images, search, or safety checking. Gemma has separate members for several of these.
• How open the maker is about training. Most publish the numbers and little else. Ai2's OLMo goes further, releasing the code, the checkpoints and the training details.

Backed by OLMo 2 7B model card, Gemma models overview.

Why a ranking goes stale

This lesson will not tell you which model is best today. Any page that claims to is out of date, and this one would be too.

New versions land often, and the terms move with them. Google's Gemma Terms of Use covers the models listed in its own appendix. For Gemma 4, that same page sends you somewhere else.

Where it sends you is worth seeing. The Gemma 4 licence page is the Apache License 2.0, the same plain licence Qwen and OpenAI use. Same family, new version, a different deal.

Notice which way that moved. Terms can loosen as well as tighten. A lesson that only ever warned you would be telling half the story.

The four differences above move much more slowly. A maker's habits, a family's range of sizes and the licence it ships under tend to hold for a good while.

So do not memorise a leaderboard. Learn the makers and the licences, then check the model's own page for what is current.

Backed by Gemma Terms of Use, Gemma 4 license, Gemma 4 model card: sizes, dense and MoE variants, active parameters, Gemma models overview.

Placing a name you have not seen

You will meet a name this course never mentions. Try the method on one now: gemma-4-26B-A4B-it.

Gemma is the family, so the maker is Google. The 4 is the version, and the version decides the terms. Gemma 4 is Apache-2.0.

26B is the total size and A4B is the active part. Memory follows the 26. Speed follows the 4.

The it on the end marks the instruction-tuned version. Its card proves it: the base model listed there is the same name without those two letters.

You have opened no benchmark, and you already know roughly whether it fits, how quick it will feel, and what you may do with it.

Three questions do that work every time.

• Who made it? Read the maker's own page, not somebody's summary of it.
• Which family, which version, and which rung? The version decides the terms, and the size tells you whether it fits.
• Total or active? If there are two numbers, the larger one is the one your memory has to hold.

Backed by gemma-4-26B-A4B-it model card and metadata, Gemma 4 license, Gemma 4 model card: sizes, dense and MoE variants, active parameters.