A model card is a short, structured document that describes one AI model: what it is for, how it was trained and evaluated, and where it falls short; a system card describes a deployed model together with the safeguards, testing and policies around it. The model card format was proposed in a 2018 research paper and is now published by most large language model providers, alongside longer system cards and, in the EU, a mandatory public summary of training content. This page explains what each document is, what the main providers actually publish, how to read one, and what a card does not tell you. It is for developers, buyers, students and governance teams who need to find and interpret these documents. Providers are listed alphabetically; the order is not a ranking.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Last checked: 2026-11-15
Key takeaways
- Model card, system card and training summary are three different documents: a model card describes the model, a system card describes testing and safeguards for a deployment, and the EU training content summary describes data sources.
- The format started in research: "Model Cards for Model Reporting" by Mitchell and colleagues was first submitted on 5 October 2018 and presented at FAT* 2019.
- Providers use the names differently: some publish a "model card" that contains frontier safety testing, others publish a "system card" for the same content, and developer model pages carry the technical specifications.
- Length varies by orders of magnitude: the documents examined for this page range from a short fields-only page to a system card whose table of contents runs past page 316.
- A card states what the provider chose to report: it is provider-reported evidence, so check the date, the exact model version it covers and any later addenda.
What is a model card?
A model card is a document that sits beside a trained model and tells a reader what the model is, what it was built for and how it performed in testing. The original paper, Model Cards for Model Reporting (Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji and Timnit Gebru), describes model cards as short documents that provide benchmarked evaluation in a variety of conditions and disclose the intended context of use and the details of the evaluation procedures.
On Hugging Face, the largest public model hub, a model card is the README.md file in a model repository: Markdown text with a YAML metadata header, rendered as the model page. Hugging Face describes cards as essential for discoverability, reproducibility and sharing, and asks authors to cover the model, intended uses and limitations (including biases and ethical considerations), training parameters, training datasets and evaluation results.
| Hugging Face metadata field | What it records |
|---|---|
license | Licence identifier; license: other with license_name and license_link for custom licences |
base_model | The model or models this one was derived from, with optional base_model_relation (for example quantized) |
datasets | Hub dataset identifiers used for training |
language | ISO 639-1 language codes |
model-index | Structured evaluation results (older format; Hugging Face documents a newer eval results format separately) |
pipeline_tag | Task type, such as text-generation, which drives search filters and widgets |
library_name | The library that loads the model |
new_version | Link to a newer version of the model |
As of 15 November 2026, from the Hugging Face model cards documentation.
What is a system card?
A system card covers a model as it is deployed: the evaluations run before release, the safeguards and usage policies applied, and the risk thresholds the provider assessed against. Where a model card answers "what is this model?", a system card answers "what did the provider test, what did it find and what did it put in place before releasing it?". System cards are usually much longer than model cards and are often updated with addenda when a new version of the same model family ships.
Frontier providers tie these documents to their own published safety frameworks. In the documents examined for this page, Google's Gemini 4 Hydrogen model card reports a Frontier Safety Framework assessment, OpenAI's GPT-6 Astra safety overview reports a Preparedness Framework level, Anthropic's Fable 5 and Mythos 5 system card has a section titled "RSP evaluations", and xAI's Grok 4.20 model card refers to its Frontier Artificial Intelligence Framework.
Model card vs system card vs training content summary
| Field | Model card | System card | EU public summary of training content |
|---|---|---|---|
| Describes | One trained model | A model as deployed, with safeguards | The data used to train a general-purpose AI model |
| Typical contents | Purpose, specifications, training summary, evaluation results, limitations, licence | Capability and risk evaluations, safeguards, alignment and misuse testing, framework thresholds | Provider and model identification, data source categories, data processing (including text and data mining opt-outs) |
| Origin | Mitchell et al., 2018 research paper | Provider practice | EU AI Act Article 53(1)(d) |
| Mandatory? | No general legal requirement | No general legal requirement | Yes, for general-purpose AI model providers placing models on the EU market |
| Typical length | One page to a few pages | Tens to hundreds of pages | A few pages in a fixed template |
As of 15 November 2026.
For how a card relates to an AI bill of materials and a dataset datasheet, see AI BOM vs Model Card vs Datasheet.
What do the main providers publish?
The table lists one recent document per provider that was read for this page. It records what each document is and how it is organised, not how good the model is.
| Provider | Document examined | Format | Date | Main sections | Safety framework referenced |
|---|---|---|---|---|---|
| Anthropic | Claude Fable 5 and Claude Mythos 5 system card | 9 June 2026 (changelog updates 11 and 25 June 2026) | Introduction; RSP evaluations; Cyber; Safeguards and harmlessness; Agentic safety; Alignment assessment; Model welfare assessment; Capabilities; Appendix | Responsible Scaling Policy (section "RSP evaluations") | |
| DeepSeek | DeepSeek-V4.1-Flash-1017 Hugging Face model card | Hugging Face README | Not stated by DeepSeek on the card | Introduction; Prompt Encoding; Minimal Inference; License; Citation; Contact | Not stated by DeepSeek |
| Gemini 4 Hydrogen model card | Web page | 12 November 2026 | Model Information; Model Data; Implementation and Sustainability; Distribution; Evaluation; Intended Usage and Limitations; Safety & Security | Frontier Safety Framework | |
| Mistral AI | Mistral Large 4 model card | Documentation page (fields, no section headings) | Release date 5 November 2026 | Fields: API names, status, type, release date, version, context length, parameters, input, output, features, pricing, licence, open weights link | Not stated by Mistral AI |
| OpenAI | Safety overview: GPT-6 Astra, linking the full system card; Deployment Safety Hub | Web page and system card | 3 September 2026 | Seven numbered points from cyber capabilities to safety for users under 18 | Preparedness Framework |
| xAI | Grok 4.20 model card | PDF, 8 pages | 7 April 2026 | Introduction; Malicious Use Risk; Loss of Control Risk; Dual-Use Capability Assessment | Frontier Artificial Intelligence Framework |
As of 15 November 2026. Each row is one example document; providers publish other cards for other models.
Anthropic
Anthropic publishes system cards as PDFs. The system card for Claude Fable 5 and Claude Mythos 5 is dated 9 June 2026, with changelog updates on 11 and 25 June 2026. It describes the two as configurations of the same underlying model: Fable 5 for general use with additional safeguards that block high-risk tasks in areas such as biology and cybersecurity, and Mythos 5 with those safeguards lifted, available only to vetted partners beginning with Project Glasswing. The table of contents places its last section on page 316.
DeepSeek
DeepSeek publishes model cards as Hugging Face README files for its open-weight models. The DeepSeek-V4.1-Flash-1017 card lists the licence as MIT, shows 763B parameters with BF16, F32, F8_E4M3 and I8 tensor types, gives a recommended context window of 1M tokens with max_tokens of at least 256K, and notes test cases for "thinking mode" and "numeric reasoning effort" in its prompt encoding section. A separate system card is not stated by DeepSeek.
Google DeepMind keeps a model cards index describing cards as simple, structured overviews of how an advanced AI model was designed and evaluated. As of 15 November 2026 it lists 30 Gemini cards plus Gemma, generative media and robotics cards, each with a published or updated date. The Gemini 4 Hydrogen card (12 November 2026) states a knowledge cutoff of March 2026 and a 1M-token context window, and reports that the model did not reach the CBRN or cyber critical capability levels but did reach their alert thresholds.
Mistral AI
Mistral AI publishes model cards as documentation pages made of fields. The Mistral Large 4 card lists API names mistral-large-4, mistral-large-4-0 and mistral-large-latest, status GA, release date 5 November 2026, version 26.10, a 1M context length, 1.05T total parameters (52B active), text and image input, reasoning and text output, and the Mistral Ecosystem Open Weight License, with open weights linked on Hugging Face. A separate system card is not stated by Mistral AI.
OpenAI
OpenAI publishes system cards and addenda on its Deployment Safety Hub, which lists recent entries such as the GPT-6 Cyber System Card (5 November 2026) and addenda to the GPT-6 Astra system card for GPT-6.1 Sol (29 September 2026) and GPT-6.1 Terra (26 October 2026). The GPT-6 Astra safety overview (3 September 2026) states that Astra is OpenAI's first model to reach the Critical level of cybersecurity capability under its Preparedness Framework. Technical specifications sit on developer model pages instead.
xAI
xAI publishes model cards as PDFs. The Grok 4.20 model card is dated 7 April 2026, runs to 8 pages and covers Grok 4.20 as deployed in single-agent and multi-agent forms. xAI also publishes training content summaries, such as the Public Summary of Training Content for Grok 4.5, last updated 8 July 2026 with an EU placement date of 14 July 2026. Technical specifications sit on developer model pages.
Where do model specifications live if not in the card?
Several providers keep the numbers developers need (context window, output limit, pricing, rate limits) on developer model pages rather than in the safety document. These pages change more often than cards, so they are where to check current values.
| Provider | Example developer page | Fields shown |
|---|---|---|
| Mistral AI | Mistral Large 4 model card (documentation) | API names, status, release date, version, context, parameters, modalities, features, pricing, licence |
| OpenAI | GPT-6.1 Sol model page | Context window (1,050,000 tokens), max output (128,000), knowledge cutoff (30 April 2026), reasoning effort levels, modalities, pricing, endpoints, snapshots, rate limits by tier |
| xAI | Grok 4.7 model page | Context window (500,000), modalities, reasoning efforts, pricing, rate limits, region, Batch API support |
As of 15 November 2026. Only pages read for this guide are listed. For how to compare context windows and prices across providers, see LLM Context Windows Explained and LLM Pricing Explained.
What is the EU public summary of training content?
Under Article 53(1)(d) of the EU AI Act, providers of general-purpose AI models must publish a summary of the content used to train the model using a template from the European Commission. The Commission FAQ states that using the template is mandatory, that providers must publish summaries from 2 August 2025, and that models placed on the market before 2 August 2025 must have summaries available no later than 2 August 2027. The template has three sections: general information, list of data sources, and relevant data processing aspects.
xAI's Grok 4.5 summary shows the template in practice: sections for provider and model identification, data source categories (publicly available datasets, licensed datasets, unlicensed private datasets, web crawling, user data, synthetic data, other sources) and data processing, including respect for text and data mining reservations and removal of illegal content. For the wider obligations, see the EU AI Act GPAI Obligations Guide.
How do you read a model card or system card?
- Check the exact model and version: confirm the card names the API identifier or version you use. A card for one release may not cover a later snapshot, and system cards are often extended by addenda.
- Check the date and changelog: note the published and updated dates. Anthropic's system card above has a changelog; Google's index shows an updated date per card; OpenAI's hub lists addenda by date.
- Separate specifications from evaluations: context, pricing and limits belong on developer pages; risk and safety findings belong in the card. Take each from the right place.
- Read benchmark tables as provider-reported: note the evaluation settings (reasoning effort, tools, number of attempts) and compare with independent sources only like for like. See How to Read LLM Benchmarks.
- Find the risk thresholds: look for the provider's framework name and the level reported, for example a Critical cyber level or an alert threshold.
- Read intended use and limitations: these sections state what the provider does and does not support, which matters for acceptable use and procurement.
- Check licence and data terms separately: for open-weight models the licence is on the card; data retention and training use are usually in separate terms and documentation.
What does a model card not tell you?
- How the model behaves in your system: a card covers the provider's tests, not your prompts, tools, data or users. Document your own deployment with the Model Card and System Card Template.
- Independent verification: results are reported by the provider. Independent leaderboards publish separately and on their own dates.
- Current prices and limits: these change faster than cards; use the developer model page and pricing page.
- Full training data: even the EU summary lists categories and main sources, not complete datasets.
- Security of your integration: keys, identities and data flows are your responsibility. See the LLM Security Guide.
Are model cards required by law?
No general law requires a document called a model card or system card. The EU AI Act requires general-purpose AI model providers to keep technical documentation and to publish the training content summary described above, and high-risk AI systems carry their own technical documentation duties. Providers publish model and system cards largely under their own safety frameworks and voluntary commitments. This section summarises public sources and is not legal advice.
Practitioner checklist
- Record the card URL, date and model version for every model in your AI inventory.
- Check for addenda or updated dates when a provider ships a new snapshot.
- Take specifications from the developer model page and pricing page, dated.
- Note the provider's safety framework and the levels its card reports.
- Keep provider-reported and independent benchmark figures in separate columns.
- For open-weight models, read the licence field and the
base_modellineage on the Hugging Face card. - For EU deployments, locate the provider's public summary of training content.
- Write your own system card for the deployed application, linking the provider card rather than copying it.
Standards and references
All sources accessed 15 November 2026.
- arXiv: Model Cards for Model Reporting (Mitchell et al.)
- European Commission: Template for general-purpose AI model providers to summarise their training content (FAQ)
- Hugging Face: Model Cards documentation
- Anthropic: Claude Fable 5 and Claude Mythos 5 System Card
- DeepSeek: DeepSeek-V4.1-Flash-1017 model card
- Google DeepMind: Gemini 4 Hydrogen model card
- Mistral AI: Mistral Large 4 model card
- OpenAI: Safety overview, GPT-6 Astra
- OpenAI: GPT-6.1 Sol model page
- xAI: Public Summary of Training Content for Grok 4.5
- xAI: Grok 4.7 model page
Related NHI Mgmt Group resources: Model Card and System Card Template, AI BOM vs Model Card vs Datasheet, EU General-Purpose AI (GPAI) Code of Practice Guide, Reasoning Models Explained.