AI // Model Catalog

Every major AI assistant, the company behind it, and whether each model is frontier (closed, cloud only), hybrid (cloud and your own hardware) or local (runs on your own machine). New to the terms? Start with the glossary.

26 companies | frontier, hybrid, local | lineups checked 13 Sep 2026

Logos and screenshots belong to their owners and are here so each product can be recognized.

App name to company

The name on the app is usually not the name of the company. This table maps one to the other, and shows what kind of model sits behind each app.

Filter

You know it as	Made by	Top model today	Kind	Good to know
ChatGPT	OpenAI	GPT-6 Astra	Frontier	Astra needs a paid plan. On Plus it sits in Work and Codex; regular chat tops out at GPT-5.6 Sol.
Claude	Anthropic	Claude Fable 5.1	Frontier	Also Opus 5, Sonnet 5 and Haiku 4.5. Claude Code is the coding tool.
Gemini	Google (DeepMind lab)	Gemini 3.1 Pro	Frontier	Called Bard until February 2024. Gemini 3.5 Pro is announced but not public yet.
Copilot	Microsoft	A mix: OpenAI GPT, Anthropic Claude, Microsoft MAI	Frontier	Built into Windows, Edge and Microsoft 365. Microsoft does not make GPT; it invests in OpenAI.
GitHub Copilot	Microsoft (owns GitHub)	Your pick: GPT, Claude, Gemini or MAI-Code-1-Flash	Frontier	Coding assistant that lives inside code editors.
Grok	SpaceXAI (was xAI)	Grok 4.6	Frontier	Elon Musk's AI, built into X (Twitter). xAI became part of SpaceX in 2026.
Meta AI	Meta	Muse Spark	Frontier	Inside WhatsApp, Instagram, Facebook, Messenger and Ray-Ban Meta glasses. Muse replaced Llama as Meta's top model.
Siri	Apple	Apple Foundation Models	Hybrid	Small requests run on the phone itself, bigger ones in Apple's cloud. Built with Google's Gemini.
Alexa+	Amazon	Amazon Nova and Anthropic Claude	Cloud only	Echo speakers and the Alexa app.
Le Chat	Mistral AI	Mistral Large 3	Hybrid	French lab. Use it in Le Chat, or download the model.
DeepSeek	DeepSeek	DeepSeek V4-Pro	Hybrid	Chinese. Free app, and the model is free to download under the MIT license.
Qwen	Alibaba	Qwen3.8-Max	Hybrid	Chinese. Comes in many sizes; the small ones run locally.
Kimi	Moonshot AI	Kimi K3	Hybrid	Chinese. 2.8 trillion parameters, billed as the largest downloadable model.
Z.ai	Zhipu AI	GLM-5.3	Hybrid	Chinese. The chat app used to be called ChatGLM.
MiniMax	MiniMax	MiniMax M3	Hybrid	Chinese. Hailuo is its video generator.
Doubao	ByteDance	Seed models	Cloud only	Made by TikTok's parent company.
ERNIE Bot	Baidu	ERNIE 5.1	Cloud only	Assistant from China's biggest search engine. The older ERNIE 4.5 is downloadable.
Perplexity	Perplexity	Sonar, plus GPT, Claude and Gemini	Cloud only	AI search engine that rents other labs' models.
Midjourney	Midjourney	Midjourney V8.1	Cloud only	Images, and short video, from a text prompt.
Sora	OpenAI	Sora	Frontier	Video generator.
Veo	Google	Veo 3.1	Frontier	Video generator with sound, inside Gemini.
Gemma, Llama, Phi	Google, Meta, Microsoft	Gemma 4, Llama, Phi	Local	No app of their own. You download them and run them offline with a tool like Ollama or LM Studio.

Frontier, hybrid or local

Every model on this page carries one of these labels. The difference is where the model actually runs, and whether you can ever get the model file.

Frontier

Closed. Cloud only.

The best models from the five frontier labs. You reach them only through the company's app or API, and the model file is never released. Needs an internet connection and usually a subscription; your prompts go to the company's servers.

Claude Fable 5.1, GPT-6 Astra, Gemini 3.1 Pro, Grok 4.6, Muse Spark

Hybrid

Both. Cloud and your own hardware.

The maker runs the model in its own app and API, and the model file is also free to download, or the work is split between your device and the company's cloud (Apple). The big ones need a multi-GPU server to run yourself, not a home PC.

DeepSeek V4-Pro, Kimi K3, Qwen3.8-Max, Mistral Large 3, Apple Foundation Models

Local

Runs on your own PC, laptop or phone.

Small downloadable models you run offline. Nothing leaves your machine and there is no usage fee, but they sit well behind the frontier. Good for private documents, offline use and tinkering.

Gemma 4, gpt-oss-20b, Qwen3.8-27B, Ministral 3, Llama, Phi, Stable Diffusion

Cloud only is used for closed models from companies outside the five frontier labs, such as Amazon Nova, Microsoft MAI, ERNIE 5.1 and Midjourney. They work the same way as frontier models (online, never downloadable) but are not at the very top.

	Frontier	Hybrid	Local
Where it runs	The company's data centers	The company's cloud, or your own server or device	Your own PC, laptop or phone
Download the model	No	Yes (Apple: built into the device)	Yes
Works offline	No	Only when you run it yourself	Yes
Privacy	Prompts go to the company	Your choice	Nothing leaves the machine
Cost	Subscription or pay per use	Free file; pay for hardware or for the maker's API	Free once you own the hardware
Capability	The best available	Close behind the frontier	Fine for everyday tasks, well behind

Running models locally

Local models are downloaded from Hugging Face or a model library and run with a free tool: Ollama (simplest, one command per model), LM Studio (desktop app with a chat window) or llama.cpp (the engine under many of these). What fits depends mostly on memory: graphics card memory on a PC, unified memory on a Mac.

Model size	Memory needed (rough, 4-bit)	Runs on	Examples	Kind
1B to 4B	2 to 4 GB	Phones and any recent laptop	Gemma 4 E2B and E4B, Ministral 3 3B, Gemini Nano	Local
7B to 14B	6 to 12 GB	A laptop or mid-range gaming GPU	Ministral 3 8B and 14B, Llama 3.1 8B, Phi	Local
20B to 32B	16 to 24 GB	A high-end gaming GPU or a 32 GB Mac	gpt-oss-20b, Gemma 4 26B and 31B, Qwen3.8-27B	Local
100B to 130B	64 to 80 GB	One data-center GPU, or a workstation with 128 GB	gpt-oss-120b, Mistral Medium 3.5	Hybrid
300B and up	Hundreds of GB	Multi-GPU servers	DeepSeek V4, Kimi K3, Qwen3.8-Max, GLM-5.3, MiniMax M3, Mistral Large 3	Hybrid

"B" is billions of parameters. For mixture-of-experts models the total size decides the memory needed, even though only part of the model works on each word.

How they connect

The companies are tangled together. These are the links that explain most of the confusion.

Who	Connected to	How
Microsoft	OpenAI	OpenAI's biggest outside investor. Copilot runs OpenAI models, and also Anthropic's Claude and Microsoft's own MAI models.
Amazon	Anthropic	Major investor. AWS is a main home for Claude, and Alexa+ runs on Amazon Nova plus Claude.
Google	Anthropic	Also an investor in Anthropic, while competing with its own Gemini.
Apple	Google	The Apple Foundation Models behind the new Siri were built with Google's Gemini (deal announced 12 Jan 2026).
SpaceX	SpaceXAI	Bought xAI on 2 Feb 2026 and renamed it SpaceXAI on 6 Jul 2026. Grok and X (Twitter) came with it.
NVIDIA	Everyone	Sells the GPU chips nearly every lab here trains and runs its models on.
Perplexity, GitHub Copilot	Many labs	Apps that let you pick another company's model instead of only building their own.

Glossary

LLM (large language model): The engine inside a chatbot. It is trained on a huge amount of text to predict what comes next, which turns out to be enough to write, code and reason.
Model vs. app: ChatGPT is the app; GPT-6 Astra is the model running inside it. One app can offer several models, and one model can power many apps. Claude, for example, runs inside Copilot and Alexa+.
Frontier model: The most capable models that exist at a given moment, from the few labs that can afford to train them. On this page the label also means closed and cloud only.
Hybrid model: A model you can use both ways: in the maker's cloud, or on your own hardware. Usually that means downloadable weights that the maker also hosts.
Local model: A model small enough to download and run on your own PC, laptop or phone, with no internet connection.
Open weights: The model file itself is published, so anyone can download it. Every hybrid and local model on this page has open weights, except Apple's models and Gemini Nano, which ship built into devices.
Reasoning model: Spends extra time working through a problem before answering. Slower and more expensive, but better at math, code and planning.
Parameters: The size of a model, counted in billions (B) or trillions (T). "1.6T total / 49B active" describes a mixture-of-experts model, which switches on only part of itself for each word.
Quantization: Storing a model's numbers with fewer bits (for example 4-bit) so it fits in less memory, at a small cost in quality. It is how 20B to 30B models fit on one gaming GPU.
Context window: How much text a model can take in at once, measured in tokens. A token is about three quarters of a word, so 1M tokens is roughly 750,000 words.
API: The paid developer connection that lets other companies build a model into their own products.

Frontier labs

The five companies building the most capable general models. Nearly everything they sell is frontier (cloud only); a few small models are downloadable.

Anthropic makes Claude

Based in: San Francisco
Founded: 2021 by Dario and Daniela Amodei and other former OpenAI researchers
Backed by: Amazon and Google

Claude
Claude Code

Model	What it is	Kind
Claude Fable 5.1Top	Anthropic's most capable model, for coding, knowledge work and long agent tasks. Released 1 Sep 2026.	Frontier
Claude Mythos 5.1	The same model as Fable 5.1 with different safeguards. Restricted to vetted cybersecurity and life-sciences programs.	Frontier
Claude Opus 5	Close to Fable at about half the price, with a low, medium or high effort setting. Released 24 Jul 2026.	Frontier
Claude Sonnet 5	Mid-size everyday model.	Frontier
Claude Haiku 4.5	Small, fast and cheapest. Still cloud only.	Frontier

OpenAI makes ChatGPT

Based in: San Francisco
Founded: 2015; CEO Sam Altman
Backed by: Microsoft, its largest outside investor

ChatGPT
Codex
Sora

Model	What it is	Kind
GPT-6 AstraTop	OpenAI's most capable model, built for computer use, browsing, coding and science. Released 3 Sep 2026. Pro, Business and Enterprise plans also get GPT-6 Astra Pro.	Frontier
GPT-5.6 Sol	The previous flagship. Still the top choice in regular chat on the Plus plan.	Frontier
GPT-Live-1	Real-time voice model that can listen and talk at the same time. Developer API, Sep 2026.	Frontier
ChatGPT Images 2.5	Image generation inside ChatGPT. Released 8 Sep 2026.	Frontier
Sora	Video generation, as its own app and model.	Frontier
gpt-oss-120b	Downloadable model (Aug 2025). Needs one 80 GB data-center GPU; also hosted by cloud providers.	Hybrid
gpt-oss-20b	Smaller downloadable model (Aug 2025). Runs on a PC or laptop with 16 GB of memory.	Local

Google DeepMind makes Gemini

Based in: London; part of Google (Alphabet)
Formed: 2023, when Google Brain and DeepMind merged; CEO Demis Hassabis
Old name: The Gemini app was called Bard until February 2024

Gemini
Gemma
Veo

Model	What it is	Kind
Gemini 3.1 ProTop	Google's top public model in the Gemini app and API.	Frontier
Gemini 3.5 Pro	Not public yet. Announced at Google I/O on 19 May 2026 and used inside Google; the release has slipped.	Frontier
Gemini 3.7 Flash	Newest fast "workhorse" model, aimed at coding and agents.	Frontier
Gemini 3.6 Flash, 3.5 Flash-Lite	Cheaper models for high-volume work (July 2026).	Frontier
Gemini 3.5 Flash Cyber	Finds and patches software vulnerabilities. Restricted to governments and selected partners.	Frontier
Veo 3.1	Video generation with sound.	Frontier
Gemma 4	Downloadable family in four sizes (E2B, E4B, 26B, 31B), Apache 2.0, released 2 Apr 2026. The small ones run on phones.	Local
Gemini Nano	Built into Android phones and Chrome; works without an internet connection.	Local

SpaceXAI makes Grok (formerly xAI)

Founded: 2023 by Elon Musk, as xAI
Owned by: SpaceX, which bought xAI on 2 Feb 2026 and renamed it SpaceXAI on 6 Jul 2026
Lives in: X (Twitter), grok.com and Tesla cars

Grok
xAI (old logo)

Model	What it is	Kind
Grok 4.6Top	Newest flagship, tuned for coding and agents, with a 500K-token context window. Released 12 Aug 2026.	Frontier
Grok 4.5	Previous flagship, released 8 Jul 2026.	Frontier
Grok 4 Heavy	Multi-agent version of Grok 4 on the most expensive plan.	Frontier
Grok 5	Reported to be in training.	Frontier
Grok Imagine	Image and video generation.	Frontier

Meta makes Meta AI

Based in: Menlo Park, California
AI lab: Meta Superintelligence Labs
Lives in: WhatsApp, Instagram, Facebook, Messenger and Ray-Ban Meta glasses

Meta AI

Model	What it is	Kind
Muse SparkTop	Meta's top model and the first from Meta Superintelligence Labs (Apr 2026), with point updates since. Closed, unlike Llama.	Frontier
Muse Image	Image generation (Jul 2026).	Frontier
Llama	Downloadable family. Llama 4 (Apr 2025) is a generation behind; the small Llama 3 sizes are still widely run on home PCs.	Local

Big tech

Giants that mostly package other labs' models into their products, while building some of their own.

Microsoft makes Copilot

Based in: Redmond, Washington
AI chief: Mustafa Suleyman, Microsoft AI
Models used: OpenAI's GPT, Anthropic's Claude, and its own MAI models

Copilot
GitHub Copilot

Model	What it is	Kind
MAI-Thinking-1Top	Microsoft's first in-house reasoning model, launched at Build in June 2026.	Cloud only
MAI-Code-1-Flash	Fast coding model rolling into GitHub Copilot.	Cloud only
MAI-Image-2.5	Image generation and editing.	Cloud only
MAI-Voice-2	Speech generation.	Cloud only
Phi	Small downloadable models that run on ordinary PCs.	Local

Apple makes Siri and Apple Intelligence

Based in: Cupertino, California
Partner: Google. The new models were built with Gemini (announced 12 Jan 2026)
Runs on: The device itself, plus Apple's Private Cloud Compute servers

Siri

Model	What it is	Kind
Apple Foundation ModelsTop	A small model runs on the iPhone, iPad or Mac itself; bigger requests go to Apple's Private Cloud Compute servers. Powers Siri AI in iOS 27, previewed at WWDC on 8 Jun 2026.	Hybrid

Amazon makes Alexa+ and Nova

Based in: Seattle
Invests in: Anthropic
Cloud: AWS Bedrock rents out Claude, Llama, Mistral and others to businesses

Alexa+
Nova
AWS

Model	What it is	Kind
Amazon Nova 2Top	Amazon's own family (Lite, Pro, Sonic) covering text, images and real-time speech.	Cloud only
Nova Act	Agent that operates a web browser for you.	Cloud only

NVIDIA makes the chips

Based in: Santa Clara, California
Why it matters: Nearly every lab in this catalog trains and runs its models on NVIDIA GPUs, and most local models run on its gaming cards

Model	What it is	Kind
Nemotron	Downloadable models in sizes from laptop-friendly Nano up to data-center Ultra. NVIDIA also hosts them.	Hybrid

China

Most Chinese labs are Hybrid: they run their own chat apps and also publish the model files. The flagships are far too big for a home PC, but their small sizes run locally. Version numbers change almost monthly.

DeepSeek

Based in: Hangzhou
Founded: 2023 by Liang Wenfeng, funded by the hedge fund High-Flyer
Known for: The R1 release in January 2025 that knocked US AI stocks

DeepSeek app

Model	What it is	Kind
DeepSeek V4-ProTop	1.6T total / 49B active parameters, MIT license. First released 24 Apr 2026; current checkpoint August 2026.	Hybrid
DeepSeek V4-Flash	284B total / 13B active. Cheaper and faster, still server-sized.	Hybrid

Alibaba makes Qwen

Based in: Hangzhou
Known for: The widest range of downloadable model sizes of any lab

Qwen

Model	What it is	Kind
Qwen3.8-MaxTop	Multimodal flagship. Its text weights were opened on 12 Aug 2026.	Hybrid
Qwen3.8-27B	Small multimodal model, Apache 2.0 license. Fits on one high-end gaming GPU.	Local

Moonshot AI makes Kimi

Based in: Beijing
Founded: 2023 by Yang Zhilin

Kimi

Model	What it is	Kind
Kimi K3Top	2.8 trillion parameters, billed as the largest downloadable model. Released 17 Jul 2026; weights published 27 Jul.	Hybrid

Zhipu AI makes GLM and Z.ai

Based in: Beijing
Founded: 2019, out of Tsinghua University
Old name: The chat app was ChatGLM; the international brand is now Z.ai

Z.ai
ChatGLM

Model	What it is	Kind
GLM-5.3Top	Current flagship, focused on coding.	Hybrid
GLM-5.3-Flash	320B total / 18B active, multimodal.	Hybrid

MiniMax makes MiniMax and Hailuo

Based in: Shanghai
Founded: 2021

MiniMax
Hailuo

Model	What it is	Kind
MiniMax M3Top	428B total / 23B active, about 1M-token context, under MiniMax's own license. Released 1 Jun 2026.	Hybrid
Hailuo	Video generation app.	Cloud only

ByteDance makes Doubao

Based in: Beijing
Also owns: TikTok and Douyin

Doubao

Model	What it is	Kind
Seed modelsTop	ByteDance's in-house family behind the Doubao chatbot.	Cloud only
Seedance 2.0	Video generation.	Cloud only

Baidu makes ERNIE

Based in: Beijing
Known for: China's largest search engine

ERNIE Bot (Wenxin)

Model	What it is	Kind
ERNIE 5.1Top	Baidu's newest model, released 8 May 2026. Hosted only; the weights are not released.	Cloud only
ERNIE 4.5	Previous generation, released as ten downloadable sizes in 2025, from 0.3B up to 424B.	Hybrid

Europe and others

Mistral AI makes Le Chat

Based in: Paris
Founded: 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix
Known for: Europe's leading lab; nearly every model is downloadable

Le Chat

Model	What it is	Kind
Mistral Large 3Top	675B total / 41B active, Apache 2.0 license. Released 2 Dec 2025.	Hybrid
Mistral Medium 3.5	128B, downloadable under a modified MIT license; self-hosts on about four GPUs. Apr 2026.	Hybrid
Ministral 3	Small models in 3B, 8B and 14B sizes, Apache 2.0. Run on laptops.	Local

Cohere makes Command

Based in: Toronto
Founded: 2019; co-founder Aidan Gomez co-wrote the 2017 Transformer paper
Sells to: Businesses and governments, not consumers

Model	What it is	Kind
Command A+Top	218B total / 25B active, reads text and images, 48 languages, Apache 2.0. Runs on two data-center GPUs. Released 20 May 2026.	Hybrid

Perplexity makes AI search

Based in: San Francisco
Founded: 2022 by Aravind Srinivas and others
Models used: Its own Sonar models, plus GPT, Claude and Gemini
Kind: Cloud only

Perplexity

Image, video and voice

Labs that make pictures, video, voices and music rather than chat.

Company	Model	Makes	Kind
Midjourney	Midjourney V8.1	Images and short video. San Francisco.	Cloud only
Black Forest Labs	FLUX.2	Images. Pro runs in the cloud; dev and klein are downloadable, and klein runs on gaming GPUs. Freiburg, Germany.	Hybrid
OpenAI	Sora, ChatGPT Images 2.5	Video and images.	Frontier
Google	Veo 3.1	Video with sound.	Frontier
SpaceXAI	Grok Imagine	Images and video.	Frontier
Runway	Gen-4.5	Video. New York.	Cloud only
Kuaishou	Kling 3.0	Video. Beijing.	Cloud only
ByteDance	Seedance 2.0	Video.	Cloud only
MiniMax	Hailuo	Video.	Cloud only
ElevenLabs	ElevenLabs voice models	Voices, dubbing and speech.	Cloud only
Suno	Suno	Songs with vocals from a text prompt.	Cloud only
Stability AI	Stable Diffusion	Images. The classic model people run at home on a gaming GPU. London.	Local