The LLM Landscape

Understanding AI models beyond the familiar chatbots

When people think of AI, they often think of ChatGPT, Gemini, or Copilot. But these are just the tip of the iceberg.

The world of Large Language Models (LLMs) includes dozens of powerful models from companies around the globe—each with different strengths, trade-offs, and ways you can access them.

Click any model on the quadrant to learn more about it.

Model Categories
Frontier Proprietary
Open Weights
Specialized / Niche
Key Terms
Large Language Model (LLM)
An AI trained on massive amounts of text to understand and generate human-like language.
Open Weights
The model's learned parameters can be downloaded, run on your own hardware, and fine-tuned (additional training) for your specific needs.
Proprietary
You can only access the model through the company's website or API. The model itself stays on their servers.
Context Window
How much text the model can "remember" at once. A 128K window can hold roughly a 300-page book.
Parameters
The number of adjustable values in the model. More parameters generally means more capable, but also more expensive to run.
Training
The process of teaching a model by exposing it to massive amounts of data. Training requires enormous computing power and can cost millions of dollars.
Inference
Using a trained model to generate responses. When you chat with an AI, you're running inference. Much cheaper than training, but still requires significant computing power.
API (Application Programming Interface)
A way for software to talk to the AI. Developers use APIs to build apps powered by AI. API access is typically priced per token—you pay for what you use.
GPT-5.2
OpenAI
Frontier Proprietary
Best-in-class general reasoning & ecosystem integration
The model behind ChatGPT. GPT-5 excels at complex reasoning, coding, writing, and multi-step problem solving. It's deeply integrated into Microsoft products and has the largest developer ecosystem.
Access API / ChatGPT
Context 128K tokens
Multimodal Text, Images, Audio
← Back to overview
Gemini 3 Pro
Google DeepMind
Frontier Proprietary
Multimodal champion with massive context window
Google's flagship model can process text, images, video, and audio natively. Its 1 million token context window means it can analyze entire codebases or books in a single conversation.
Access API / Gemini App
Context 1M tokens
Multimodal Text, Images, Video, Audio
← Back to overview
Opus 4.5
Anthropic
Frontier Proprietary
Nuanced writing & reliable agentic coding
Known for thoughtful, nuanced responses and strong safety alignment. Claude excels at long-form writing, complex analysis, and working as an autonomous coding assistant that can use tools and browse the web.
Access API / Claude.ai
Context 200K tokens
Multimodal Text, Images, PDFs
← Back to overview
Grok 4.1
xAI
Frontier Proprietary
Real-time world knowledge via X integration
Created by Elon Musk's xAI, Grok has unique access to real-time data from X (formerly Twitter). This lets it answer questions about current events and trending topics that other models can't access.
Access API / X / Grok.com
Context 2M tokens (Fast)
Multimodal Text, Images, Voice
← Back to overview
Microsoft Copilot
Microsoft
Frontier Proprietary
Partnership model: integrates others' AI into Microsoft products
Unlike the others on this chart, Microsoft doesn't build its own large language models. Instead, Copilot is powered by partnerships—primarily with OpenAI (GPT models). This lets Microsoft embed AI across Office, Windows, Edge, and GitHub without the massive R&D investment of training frontier models.
Powered By OpenAI GPT models
Access Windows, Office, Edge, Bing
Strategy Partnership, not training
Small Models Phi series (in-house)
← Back to overview
Llama 4 Maverick
Meta
Open Weights
The industry standard for local deployment
Meta's open-weight model that you can download and run yourself. Maverick uses a "mixture of experts" architecture—only activating 17B of its 400B parameters per query, making it efficient yet powerful.
Access Download / Self-host
Parameters 17B active / 400B total
Context 1M tokens
License Llama Community
← Back to overview
DeepSeek V3.2
DeepSeek (China)
Open Weights
Massive efficiency with frontier-level reasoning
A Chinese model that shocked the industry by matching GPT-4 level performance at a fraction of the training cost. Released under the permissive MIT license, it's become a favorite for researchers and companies seeking cost-effective AI.
Access Download / Self-host
Parameters 37B active / 671B total
Context 128K tokens
License MIT (fully open)
← Back to overview
Qwen 3
Alibaba (China)
Open Weights
Dominant in math, coding & multilingual tasks
Alibaba's Qwen family has become the most downloaded open model in the world. Trained on 36 trillion tokens across 119 languages, it excels at mathematical reasoning and code generation.
Access Download / Self-host
Parameters 22B active / 235B total
Languages 119 languages
License Apache 2.0 (open)
← Back to overview
GPT-OSS
OpenAI
Open Weights
OpenAI's first open weights since 2019—reasoning-focused
In August 2025, OpenAI surprised everyone by releasing open-weight models for the first time since GPT-2. Available under the permissive Apache 2.0 license, these reasoning-focused models can run on a single GPU. A major strategic shift after years of being called "ClosedAI."
Access Download / Self-host
Sizes 120B and 20B variants
Hardware 1 GPU (80GB) or 16GB RAM
License Apache 2.0 (fully open)
← Back to overview
Gemma 3
Google DeepMind
Open Weights
Google's open model—multimodal with 128K context
Google's lightweight open models built from the same technology as Gemini. Gemma 3 supports images and text input, handles 140+ languages, and can run on everything from phones to gaming GPUs. Available in sizes from 270M to 27B parameters.
Access Download / Self-host
Sizes 270M to 27B parameters
Context 128K tokens
Features Multimodal, 140+ languages
← Back to overview
Mistral Small 3
Mistral AI (France)
Specialized
Extreme speed for low-latency applications
A French company's speed-optimized model that generates 150 tokens per second. Small enough to run on a single GPU or a MacBook with 32GB RAM, yet competitive with models 3x its size.
Access Download / Self-host
Parameters 24B
Speed ~150 tokens/sec
License Apache 2.0 (open)
← Back to overview
Phi-4
Microsoft
Specialized
High-end reasoning in a tiny, mobile-friendly package
Microsoft's "small language model" proves that bigger isn't always better. At just 3.8B parameters, Phi-4 rivals much larger models on math and reasoning—small enough to run on a phone.
Access Download / Self-host
Parameters 3.8B - 14B
Best For Math, Reasoning, Mobile
License MIT (fully open)
← Back to overview
Command R+
Cohere
Specialized
Optimized specifically for Enterprise RAG
Built for businesses that need AI to answer questions from their own documents. Command R+ specializes in "retrieval-augmented generation"—finding relevant info from databases and citing its sources.
Access API only
Context 128K tokens
Best For Enterprise search & RAG
Languages 10 business languages
← Back to overview
Perplexity
Perplexity AI
Specialized
AI-powered search engine with citations
Perplexity isn't a model—it's a search product built on top of multiple models (including Claude and GPT). Their innovation is combining AI with real-time web search, always citing sources. Think of it as "Google meets ChatGPT."
Type Product, not a model
Powered By Claude, GPT, + own models
Best For Research & fact-finding
Key Feature Always cites sources
← Back to overview

Early 2026 Model Landscape

Position reflects access type (open vs. proprietary) and purpose (general vs. specialized)

Open Weights
Run it yourself
Proprietary
API access only
General Purpose
Specialized
Open & General
Proprietary & General
Open & Specialized
Proprietary & Specialized
GPT-5.2
Gemini 3
Opus 4.5
Grok 4.1
Copilot
Llama 4
DeepSeek V3
Qwen 3
GPT-OSS
Gemma 3
Mistral Small
Phi-4
Command R+
Perplexity
👆 Click a model to learn more