llm-metadata: Lightweight LLM Metadata Static API, Making Model Discovery Simple

In AI application development, LLM metadata (such as model name, context length, feature support, pricing, etc.) is scattered across various providers' documents, with inconsistent formats and frequent updates. The llm-metadata project provides a fully static API service, deployable via GitHub Pages or Cloudflare Pages, with advantages like zero operation and maintenance costs, high throughput performance, multilingual localization support, and NewAPI format compatibility, making LLM metadata acquisition simple and reliable.

In AI application development, accurately obtaining LLM metadata is a common need, but this data is scattered across different providers' documents, with inconsistent formats and frequent updates, causing considerable trouble for developers. The llm-metadata project was born to solve this problem.

llm-metadata takes the "static-first" as its core philosophy, providing services through pre-built static JSON files, which can be deployed on GitHub Pages or Cloudflare Pages. This design brings four major advantages: zero operation and maintenance costs (no need to maintain servers and databases), extremely high performance (global CDN caching, millisecond-level response), unlimited scalability (no concurrency limits), and free hosting (using platform free quotas). When data changes, only static files need to be regenerated without affecting the service.

Data Sources: Integrates models.dev/api.json and community contributions (accepts corrections and supplements via the data/overrides directory). Covered Information: Basic model information, capability tags (tool calling, open-source weights, etc.), modality support (text/image/audio), context limits, pricing, provider configurations, etc. Multilingual Support: Adopts structured translation management. Adding a new language only requires three steps (update locales.json, create translation files, run build). After building, the localized API is located in the dist/api/i18n// directory. NewAPI Compatibility: Provides NewAPI format output, supporting direct use by existing toolchains.

Direct Use: Access the GitHub Pages (https://basellm.github.io/llm-metadata/) or Cloudflare Pages (https://llm-metadata.pages.dev/) hosted versions. Local Build: Install dependencies (npm install), run build (npm run build), force build (npm run build:force), CI check (npm run check). Custom Overrides: Customize data via the data/overrides directory. Override files use a deep merge strategy, supporting field modifications at the model and provider levels (e.g., id, name, description, pricing, etc.).

Automated CI/CD: Implemented via GitHub Actions. Trigger conditions include directory pushes, manual triggers, and scheduled runs every 6 hours. The incremental update strategy is controlled by data/policy.json; when auto=false, manually customized content is protected. Application Scenarios: 1. Model selector UI (filter by capability/price/context); 2. Cost estimation tool (pre-calculate usage costs); 3. Capability detection (avoid runtime errors); 4. Document generation (automatically sync the latest data).

Architectural Advantages: Demonstrates an effective model of "static generation + CDN distribution + incremental updates + community-driven", suitable for scenarios with more reads than writes and low-frequency data changes (e.g., LLM metadata). Industry Insights: With the development of the AI ecosystem, similar static-first services may emerge, such as prompt template libraries, evaluation benchmark results, model performance comparisons, etc. These services require high availability, low latency, and zero operation and maintenance, and llm-metadata provides a replicable model.

llm-metadata is a project with a concise design but practical functions. It solves the pain points of LLM metadata acquisition through static APIs, embodying the power of the "static-first" philosophy. For developers, it provides a reliable data source; for engineers, it is a case study on how to solve practical problems with minimal complexity.