Section 1

AI System Cards
& Model Cards

Public documentation published by AI providers detailing the capabilities, limitations, safety evaluations, and intended uses of their large language models.

What is a Model Card?

A model card is a short document accompanying a trained machine learning model that describes the model's intended use cases, training data, performance benchmarks, limitations, and ethical considerations. The concept was formalised by Mitchell et al. (2019) at Google.

Model cards are intended to allow practitioners, regulators, and end-users to make informed decisions about deploying or interacting with AI systems.

What is a System Card?

A system card goes a step further — it documents not just the underlying model, but the entire AI system as deployed, including fine-tuning, RLHF, safety mitigations, red-teaming results, and deployment constraints. OpenAI pioneered this format for GPT-4.

System cards are often more relevant from a regulatory standpoint because they reflect what users actually interact with, not just the base model.

EU AI Act relevance

Under the EU AI Act, providers of General-Purpose AI (GPAI) models with systemic risk are required to publish technical documentation that overlaps significantly with the concept of a system or model card. See the EU AI Act page for more detail.

Leading LLMs & Their Documentation

Click "System Card" or "Model Card" to access the official documentation published by each provider.

OpenAI

GPT-4o

OpenAI's flagship multimodal model, capable of processing and generating text, images, and audio in a single model. GPT-4o is the model powering ChatGPT Plus and the OpenAI API.

OpenAI

o3 & o4-mini

OpenAI's reasoning-focused model series, trained with reinforcement learning to "think" through problems step by step before answering. Excels at maths, science, and coding benchmarks.

Anthropic

Claude (Claude 4 Opus / Sonnet)

Anthropic's Claude family is built around Constitutional AI — a technique that trains models using a set of principles rather than purely human feedback. Claude is known for safety, nuance, and long-context capabilities.

Google DeepMind

Gemini 2.0 / 1.5 Pro

Google's Gemini family spans Flash, Pro, and Ultra tiers. Gemini 1.5 Pro introduced a 1M+ token context window. Gemini powers Google Search AI Mode, Workspace, and the Gemini API.

Meta AI

Llama 3.3 / 4

Meta's open-weights model family, released under a community licence. Llama 3 models range from 8B to 405B parameters. Llama 4 introduces multimodal capabilities and a Mixture-of-Experts architecture.

Mistral AI

Mistral Large / Mixtral 8x22B

Mistral AI's flagship proprietary and open-weights models. Mixtral uses a Sparse Mixture-of-Experts architecture, activating only a subset of parameters per token for efficient inference at high quality.

xAI

Grok-3

xAI's large language model, integrated into X (formerly Twitter). Grok-3 claims strong reasoning and coding performance, with real-time knowledge via its integration with X posts.

Cohere

Command R+

Cohere's enterprise-focused model, optimised for Retrieval-Augmented Generation (RAG) and tool use. Command R+ is designed for complex business workflows requiring reliable grounding and citations.

Technology Innovation Institute (TII)

Falcon 180B

Developed by TII in Abu Dhabi, Falcon 180B was among the largest open-access language models at its release. Licensed for commercial use, it achieved top open-source benchmarks upon release.

Amazon Web Services

Amazon Nova / Titan

AWS's family of foundation models available via Amazon Bedrock. Amazon Nova (successor to Titan) spans Micro, Lite, Pro, and Premier tiers, covering text, multimodal, and creative tasks.

IBM

IBM Granite

IBM's enterprise-grade foundation model family built for business use cases including legal, finance, and code. Granite models are published with detailed model cards covering training data transparency.

DeepSeek

DeepSeek-R1 / V3

High-performance open-weights models from Chinese lab DeepSeek AI, notable for matching frontier model performance at significantly reduced training cost. R1 is a reasoning-focused variant.

Quick Comparison

Summary of model documentation availability and licence type.

Model Provider System / Model Card Weights Access Licence
GPT-4o OpenAI System Card ✓ Closed API Proprietary
o3 / o4-mini OpenAI System Card ✓ Closed API Proprietary
Claude 4 Opus Anthropic Model Safety Card ✓ Closed API Proprietary
Gemini 2.0 Google DeepMind Technical Report ✓ Closed API Proprietary
Llama 3.3 / 4 Meta AI Model Card ✓ Open Weights Llama Community Licence
Mixtral 8x22B Mistral AI Model Card ✓ Open Weights Apache 2.0
Grok-3 xAI Limited ∂ Closed API Proprietary
Command R+ Cohere Model Card ✓ Open Weights (HF) CC-BY-NC 4.0
Falcon 180B TII Model Card ✓ Open Weights Falcon Licence
DeepSeek-R1 DeepSeek AI Model Card ✓ Open Weights MIT
Granite 3.x IBM Model Card ✓ Open Weights Apache 2.0