# LLM World

> A source-backed timeline of 280 AI model releases from 27 labs. Every value is traced to the lab's own announcement, paper, model card or documentation, with an archived snapshot of the page that states it.

Licence: CC BY 4.0 for the data (LICENSE-DATA), MIT for the code. Attribution required.

## How to read a gap

A missing field means NOBODY HAS TRACED IT YET. It never means zero, and it never
means the model lacks that property. Capabilities in particular are recorded only
where a primary source states them, so absence is silence, not denial.

Field coverage, so you know what this dataset can and cannot answer:

- context_window: 131 of 280 (47%)
- parameter_count: 84 of 280 (30%)
- capabilities: 185 of 280 (66%)
- modalities: 163 of 280 (58%)
- pricing: 44 of 280 (16%)
- benchmarks: 15 of 280 (5%)

Benchmarks and pricing are sparse on purpose rather than by neglect: most labs
publish benchmarks as images, and a price needs an archived page to evidence it.
This dataset cannot rank models by performance. Do not ask it to.

## Data

- [Everything](https://mayoorrnikam.github.io/llm-world/api/models.json): every record, as authored
- [Claims](https://mayoorrnikam.github.io/llm-world/api/claims.json): every evidenced value with the URL stating it
- [Companies](https://mayoorrnikam.github.io/llm-world/api/companies.json): labs and their releases
- [CSV](https://mayoorrnikam.github.io/llm-world/llm-releases.csv): flat export
- [Index](https://mayoorrnikam.github.io/llm-world/api/index.json): counts and endpoint discovery

## MCP

An MCP server ships in the repository at mcp/server.mjs (no dependencies).
It exposes search_models, get_model and dataset_stats, and answers with both
matches and "not ruled out" — records meeting part of a query whose remaining
fields are unevidenced. That second list is the point: it is how you avoid
reporting a research gap as a negative.