AI Model Families: Who Makes GPT, Claude, Gemini, Llama and More

ChatGPT, Claude, Gemini, Llama, DeepSeek. If AI news feels like an endless parade of names, you’re not imagining it. New models come out almost every month.

The good news is that the names fall into families. Each family comes from one company or lab, and each lab releases new versions over time. On the Periodic Table of AI, these are the 15 pink tiles in the row pulled out underneath the main chart, numbers 57 through 71.

Three glowing pink cubes: a sealed cube, a cube with its lid open showing a bright core, and a clear glass cube showing its core and streams of light flowing in
Closed, open-weight and fully open: the biggest difference between AI model families.

The last post was about prompting, or how to talk to a model. This one is about who’s on the other end.

First, one idea that sorts them all

The biggest difference between model families isn’t who’s smartest this month. That changes constantly. The lasting difference is how open each one is.

Closed models live on the company’s servers. You use them through an app or a paid connection, but you can’t download them.

Open-weight models let anyone download the finished model, meaning its parameters, and run it on their own hardware. That matters for privacy, cost and control. Licenses vary, though. Some come with strings attached, which is why many experts avoid calling them “open source.”

Fully open models go further. They also publish the training data and the code used to build the model, so anyone can see exactly how it was made.

Many companies mix approaches, releasing some models openly while keeping their best ones closed.

The big closed families

GPT

GPT is OpenAI’s family, and it’s the one that started the craze. OpenAI launched ChatGPT on November 30, 2022, and the name says what’s inside: Chat plus GPT, short for Generative Pre-trained Transformer. GPT models power ChatGPT and many other apps built on OpenAI’s service.

Claude

Claude is Anthropic’s family. It’s available through the Claude app and through Anthropic’s service for developers, and it’s a common choice for writing, analysis and coding work.

Gemini

Gemini is Google’s family of closed models. It’s built into Google’s own products and is also available to developers through Google’s cloud.

Grok

Grok comes from xAI, the company Elon Musk launched in 2023. It debuted that November with real-time access to posts on X, and it was designed with a witty, rebellious personality. That access to live posts is still one of the things that sets it apart.

Nova

Nova is Amazon’s family, announced at Amazon’s re:Invent conference in December 2024. It lives on Amazon Bedrock, Amazon’s service for businesses that want to build with AI. Alongside the text models, the original lineup included Nova Canvas for images and Nova Reel for video.

Command

Command comes from Cohere, a company founded in 2019 by former Google Brain researchers. Cohere aims squarely at businesses rather than everyday users. Its Command models are tuned for jobs like answering questions from company documents and using tools.

The open-weight families

Llama

Llama is Meta’s family, and it kicked off the open-weight movement. The first version came out in February 2023. At first, its weights were only available to approved researchers, but later versions opened up to the public and allowed many commercial uses. Llama comes with Meta’s own license, which carries some restrictions, so it’s best described as open-weight rather than open source.

Gemma

Gemma is Google’s open-weight family, released in February 2024. Google built it from the same research behind Gemini, but made the models small enough to run on a regular laptop or desktop. So Google is in both camps: Gemini is closed, and Gemma is open.

Phi

Phi comes from Microsoft Research, and its whole point is being small. Phi models are trained heavily on carefully chosen, textbook-style material, on the bet that better data can beat bigger size. When Microsoft released Phi-3 in April 2024, it said the smallest version could fit on a smartphone and work without an internet connection.

Mistral

Mistral is the family from Mistral AI, a Paris company founded in April 2023 by researchers who came from Google DeepMind and Meta. It’s Europe’s best-known AI lab. Mistral releases many of its models with open weights and sells others through its own service.

DeepSeek

DeepSeek is a lab based in Hangzhou, China. It shook up the AI world in January 2025 with R1, a reasoning model that matched OpenAI’s best reasoning model on key math benchmarks. DeepSeek released R1’s weights under the MIT license, one of the most permissive licenses there is.

Qwen

Qwen is Alibaba Cloud’s family. Most Qwen models are released as open weights, many under the permissive Apache 2.0 license, while Alibaba keeps its most advanced models proprietary. Qwen comes in a huge range of sizes and covers text, code, images and audio.

Kimi

Kimi comes from Moonshot AI, based in Beijing. Moonshot began releasing open-weight models with Kimi K2 in July 2025. In July 2026, it unveiled Kimi K3, a model with 2.8 trillion parameters that Moonshot calls the world’s largest open-weight model.

GLM

GLM is the family from Zhipu AI, a Beijing company with roots at Tsinghua University that now goes by Z.ai internationally. GLM stands for General Language Model. Since 2025, Z.ai has released many of its models under the MIT license.

The fully open family

OLMo

OLMo comes from the Allen Institute for AI, or Ai2, a nonprofit research institute. When it launched in February 2024, Ai2 released more than the finished model. It also published the full training data, the training code, the training logs and checkpoints saved during training.

That makes OLMo a rare case where anyone can see exactly what went into a model and repeat the work. Most “open” models only give you the finished product.

How to make sense of it all

Here’s the short version. If you just want a chatbot, the closed families like GPT, Claude, Gemini and Grok are the easiest to use, since you just open the app. If you want to run a model on your own computer or keep your data in-house, the open-weight families like Llama, Gemma, Qwen and DeepSeek are where to look. If you want to study how a model was built, OLMo is the one to watch.

Don’t get too attached to rankings. The lead changes hands often, and the gap between open and closed models has been shrinking.

Where to go next

Models are only part of the picture. The next group on the chart is Tools & Protocols, which covers how AI connects to other software, the web and your own data.

In the meantime, click any of the pink tiles and read the full page. Start with Llama if you want to understand why open weights became such a big deal.

1 thought on “AI Model Families: Who Makes GPT, Claude, Gemini, Llama and More”

Comments are closed.