normalizeOpenWeights

Catalog metadata is reported per provider entry, and entries for the same physical model can disagree. A model whose weights are on HuggingFace may have openWeights: true on its huggingface entry and false on reseller entries. Any split into open-weight and hosted lists then depends on which entry wins.

normalizeOpenWeights() resolves the fact per model: if any entry for a model reports openWeights: true, every entry for that model gets true. Weights are either public or they are not.

function normalizeOpenWeights<T extends Model>(models: T[]): T[]

Model identity is resolved the same way as matchesModel(). Entries of models with no true report are returned unchanged (including undefined, which stays unknown rather than becoming false).

import { fromModelsDev, normalizeOpenWeights } from "pickai";

const models = normalizeOpenWeights(await fromModelsDev());

const openWeight = models.filter((m) => m.openWeights);
const hosted = models.filter((m) => m.openWeights === false);
// The same physical model can no longer land in both lists.
// Entries with openWeights undefined are unknown, not hosted; handle
// them separately if your split needs to be exhaustive.

Run it once, right after loading the catalog, before any filtering or splitting that reads openWeights.