Track frontier AI, from original reports to technical insight.
Website · Chinese · Browse publishers · Browse document types · Frontier models · Summaries & figures · Full catalog
An open research archive for understanding how frontier AI models are built, evaluated, and deployed. We bring official technical reports, model cards, and system cards into one place, and are building toward readable sources, grounded technical analysis, and ongoing release tracking.
385 original PDFs · 18 publishers · 2022 onward
Explore the website to search reports, filter by publisher and document type, and browse current frontier models in English or Chinese.
Available today: original PDFs, a searchable bilingual website with source reading pages and figure previews, and frontier report summaries with core figures. Full-document Markdown, structured analysis, and automated tracking are the next stages.
Star AI Frontier to keep the sources close and follow the archive as it grows.
Looking for implementation details, model boundaries, or deployment safety? Start with the corresponding document type.
| Type | Main subject | Focus | Typical contents | PDFs |
|---|---|---|---|---|
| Technical Report | Model / training method | Implementation and experiments | Architecture, data, training, post-training, inference, benchmarks | 268 |
| Model Card | Individual model | Model description and usage boundaries | Capabilities, limitations, intended and unsuitable uses, evaluations, safety information | 59 |
| System Card | Complete product / system | Risk, safety, and deployment behavior | Red teaming, dangerous capability evaluations, jailbreaks, safety mitigations, deployment restrictions | 58 |
These are typical distinctions, not rigid boundaries. Documents can overlap or cover a model family; classification follows the publisher's designation and the document's primary purpose.
Pick a lab to find its archived reports and official source links. Each publisher has separate Technical Report, Model Card, and System Card folders. Each directory is ordered by recorded first publication date, newest first.
Click a count to open that publisher's collection for the document type. Counts include archived PDFs only.
| Publisher | Technical Reports | Model Cards | System Cards | Total PDFs |
|---|---|---|---|---|
| OpenAI | 6 | 1 | 35 | 42 |
| Anthropic | 0 | 5 | 17 | 22 |
| Google / DeepMind | 39 | 32 | 0 | 71 |
| Meta | 28 | 0 | 2 | 30 |
| xAI | 0 | 7 | 1 | 8 |
| Mistral AI | 11 | 0 | 0 | 11 |
| NVIDIA | 24 | 0 | 0 | 24 |
| DeepSeek | 23 | 2 | 0 | 25 |
| Alibaba / Qwen / Wan | 40 | 0 | 0 | 40 |
| Moonshot AI / Kimi | 10 | 0 | 0 | 10 |
| Z.ai / Zhipu AI | 18 | 0 | 0 | 18 |
| MiniMax | 6 | 0 | 0 | 6 |
| StepFun | 18 | 0 | 0 | 18 |
| Tencent / Hunyuan | 15 | 0 | 0 | 15 |
| ByteDance / Seed | 9 | 4 | 0 | 13 |
| Cohere | 8 | 0 | 0 | 8 |
| Microsoft / Phi / MAI | 9 | 8 | 0 | 17 |
| Amazon / Nova | 4 | 0 | 3 | 7 |
- Go straight to the source. Original publisher PDFs and official links, together in one index.
- Compare across labs. Find training methods, evaluation results, and safety disclosures across technical reports and model cards.
- Trace what changed. Publication dates and revisions are recorded separately; older archived versions are retained when reports are updated.
The full catalog includes dates, versions, sources, and file verification details. Known omissions and uncertain dates are listed in coverage and gaps (currently maintained in Chinese). Publisher and document-type directories have separate English and Chinese READMEs; report titles and original PDFs retain their source language.
Found a missing report, a newer revision, or a broken link? Open an issue or pull request with the report title, publisher, and official source. A single good source link helps make the collection more useful for everyone.
If this archive saves you a search, give it a Star or share it with someone reading AI papers.
The original reports are the foundation. The longer-term goal is a research workflow that helps readers follow releases, inspect technical evidence, and understand what changed.
| Stage | What it helps you do | Status |
|---|---|---|
| Original sources | Find official reports, source links, dates, and archived revisions | Available |
| Readable sources | Read source summaries and inspect core figures; full-document Markdown is next | Summaries and figures available |
| Technical analysis | Examine architecture, data, training, post-training, inference, and evaluation details with source references | Planned |
| Continuous tracking | Follow new model releases and report revisions through automated monitoring and update histories | Planned |
The bilingual website makes the archive searchable. Sources and frontier selections are curated manually.
This growing collection covers official model and model-family reports from 2022 onward, including technical reports, model cards, and system cards. Sources are publisher websites, official repositories, and author-submitted arXiv papers. Coverage is not yet exhaustive. For arXiv entries, the recorded first publication date is the first submission date; unknown dates are left unspecified.
PDFs are preserved as provided by their publishers. Copyrights and terms of use remain with the original rights holders; this repository does not relicense the reports.
Maintenance and implementation notes: agent.md.