Overview

The public index API: what it serves, what it costs, and where to start.

• Quickstart · Three requests that answer the three questions most callers arrive with.
• Authentication & CORS · Every read is anonymous. What that means for keys, browsers and headers.
• Caching & conditional requests · Two cache classes, what ETag means here, and how long a publish takes to reach you.
• Errors & status codes · One error envelope, ten codes, and the ones a public caller can actually see.
• Filtering, sorting & pagination · The /v1/models filter surface, its validation rules and its sort semantics.
• Concepts · Snapshots, supersession, trust, coverage, and the two different VRAM numbers.
• Models · Filter the fleet, or read one model in full — parameters, VRAM, pricing, memory profiles, throughput and every benchmark claim.
• Leaderboard · Every ranked model against every weighted core benchmark, with cost and VRAM, in a single cached payload.
• Benchmarks · The registry behind every score: units, anchors, composite weight, lifecycle status and per-benchmark coverage.
• Families & hardware · The two registries the rest of the API refers to by id: model lines with their live heads, and the hardware every throughput figure is measured on.
• Snapshots · The published body, the history behind it, and the model-level diff for one publish.
• VRAM · Weights-footprint estimates for the current fleet, and each model’s efficiency against the fleet’s own fitted frontier.
• Runs & feed · Pipeline run history, one run with its events, and the global event tail with a forward cursor.
• Health · Which snapshot is serving, how old it is, and when the data behind it last moved.