Use the data

Endpoints

/api/index.json	Discovery: licence, counts, endpoint list. Stable — point a badge at it.
/api/models.json	Every release as authored, with sources and provenance.
/api/claims.json	375 claims, denormalised: value, the lab page asserting it, the archived snapshot.
/api/companies.json	Per-lab counts, open-weights share, first and latest release.
/llm-releases.csv	Flat table for spreadsheets.
/llms.txt	What an agent should read first.

Every payload carries its own license, attribution and schema_version, so the terms travel with the data. A breaking shape change bumps schema_version — pin against it.

Licence

The data is CC BY 4.0: use it, change it, sell it, as long as you credit LLM World and link back. The code is MIT. See LICENSE-DATA and NOTICE for exactly what each covers.

Release dates and metadata from LLM World
https://mayoorrnikam.github.io/llm-world/ — CC BY 4.0

What a missing field means

Nobody has traced it yet. Not zero, and not that the model lacks the property. Capabilities are recorded only where a primary source states them, so absence is silence rather than denial — a model with no coding capability may well code, and this dataset simply has not evidenced it. Building a "does not support" claim on a blank field is the one misuse that will make you wrong.

Records	280, of which 185 (66%) are verified — every value traced to a primary source
modalities	163 of 280 (58%) — What the model takes in and puts out
capabilities	185 of 280 (66%) — Only where a primary source states them
context_window	131 of 280 (47%) — Language models only
parameter_count	84 of 280 (30%) — Most proprietary labs publish none
pricing	44 of 280 (16%) — Needs an archived page to evidence it
benchmarks	15 of 280 (5%) — Most labs publish these as images

pricing and benchmarks are thin on purpose rather than by neglect, and the reasons are in data quality. What follows from it: this dataset cannot rank models by performance or cost across the whole set. It can tell you what a lab stated and where.

MCP server

A dependency-free server over the same data, for Claude Desktop, Claude Code or any MCP client. It lives in the repository at mcp/server.mjs.

git clone https://github.com/mayoorrnikam/llm-world.git
npm run mcp

Then add it to your client's config — for Claude Desktop that is ~/Library/Application Support/Claude/claude_desktop_config.json:

{ "mcpServers": { "llm-world": {
    "command": "/opt/homebrew/bin/node",
    "args": ["/absolute/path/to/llm-world/mcp/server.mjs"] } } }

Use an absolute path to node, not "node". A GUI application on macOS inherits /usr/bin:/bin:/usr/sbin:/sbin and never the PATH from your shell profile, so a Node installed by nvm or a version manager is invisible to it and the server fails to spawn with nothing useful in the log. which node in a terminal answers about a different PATH than the one your client has.

search_models	A plain-language query. Returns matches and not_ruled_out.
get_model	One record in full, every claim with its primary and archived URL.
dataset_stats	Coverage per field, so a caller knows what the data supports.

not_ruled_out is the part worth having. Asked for a coding model with a 1M context window it returns the models that match — and separately lists Claude Sonnet 5 and Claude Fable 5, which have the context window and whose coding ability nobody has evidenced. Reporting those as "no" would be false; dropping them silently is how a research gap turns into a wrong answer.

Nothing in it ranks models. Benchmarks cover 5% of records, and a leaderboard on that basis would be the one part of this project that is not evidence.