What's My Model?

A catalog of viable AI models for your hardware.

The core is pure JavaScript with no DOM, framework, or backend dependency: given a hardware profile ({ gpu, ram } in bytes) and a model's file size, it classifies fit (ok / tight / over). Hardware detection is not baked in — each host injects it through a HardwareProvider. A browser best-effort provider ships in-package (no backend); desktop hosts supply an exact native probe. That injection seam is what lets the widget drop into any tool without carrying its own server.

Packages

https://www.npmjs.com/org/whats-my-model

• @whats-my-model/core — pure fit engine (classifyModel, context/KV-cache-aware estimateFit), GGUF discovery, ranking (recommend), and the hardware/catalog provider contracts. Zero dependencies.
• @whats-my-model/widget — the <whats-my-model> Web Component (vanilla, no build). Task / preference / context / KV-cache controls, editable hardware, a wmm-select event.
• @whats-my-model/catalog-huggingface — backend-free Hugging Face catalog: live search, a bundled snapshot, a hybrid (snapshot ∪ live), and an IndexedDB cache.
• @whats-my-model/react — a thin React wrapper for the Web Component.

Quick start

Drop the widget into any page buildlessly via an import map — see apps/demo. Or wire it in code:

import "@whats-my-model/widget";
import {
  huggingFaceCatalogProvider,
  hybridCatalogProvider,
  cachedCatalogProvider,
} from "@whats-my-model/catalog-huggingface";
// The bundled snapshot ships behind a subpath so live-only consumers don't load it.
import { snapshotCatalogProvider } from "@whats-my-model/catalog-huggingface/snapshot";

const el = document.querySelector("whats-my-model");
el.configure({
  // Hosts inject hardware — browsers can't read VRAM. Web: browserHardwareProvider
  // (coarse) or manual entry; desktop (Tauri/Electron): an exact native probe.
  hardwareProvider: { inspect: async () => ({ gpu: { total: 16 * 2 ** 30 }, ram: { total: 32 * 2 ** 30 } }) },
  // Bundled snapshot for instant/offline, folded with a cached live search.
  catalogProvider: hybridCatalogProvider(
    snapshotCatalogProvider(),
    cachedCatalogProvider(huggingFaceCatalogProvider({ task: "code" })),
  ),
  workload: { task: "code", preference: "balanced", targetContext: 32768, cacheType: "q4_0" },
});
el.addEventListener("wmm-select", (e) => console.log(e.detail.variant));

React hosts: import { WhatsMyModel } from "@whats-my-model/react" — see examples/react.

Dev: python scripts/serve.py serves the demos (no-store, so ES-module edits are picked up); node scripts/build-catalog.mjs regenerates the snapshot; node --test packages/*/src/*.test.js runs the suite.

Staying current

Two independent mechanisms keep the catalog relevant, so it never rots:

1. Runtime (live site / any host). The hybrid provider folds a live Hugging Face search into the bundled snapshot on every load — the site is current even between refreshes, and unknown models resolve on drill-in.
2. Scheduled refresh (GitHub Action, weekly). .github/workflows/data-refresh.yml runs scripts/build-catalog.mjs every Monday (and on demand) and commits the rebuilt snapshot to main only when the model data actually changed. GitHub Pages redeploys automatically after each refresh.

build-catalog.mjs is curated, not a raw popularity dump — HF's download ranking is full of no-name "…-Claude-Opus-Reasoning-Distilled / …-Uncensored" reposts. So it:

• pulls across workload + family queries (coder, instruct, reasoning, gemma, phi, mistral, llama, deepseek), restricted to trusted publishers (model authors + canonical quantizers: bartowski, unsloth, lmstudio-community, …);
• collapses each model to one publisher and a representative quant ladder (Q2_K…Q8_0 + IQ4_XS), keeping the snapshot to ~50 distinct, current families;
• self-guards: an empty fetch fails loudly and a big shrink is refused — a bad or rate-limited run keeps the last-good snapshot instead of shipping a gutted one.

The one hand-maintained knob is the trusted-publisher list at the top of the script. The refresh does not auto-publish to npm — that stays a deliberate release (Cut release), so npm consumers get catalog updates on the next version you cut.

Status

v0.1 — the fit engine, extracted from Concurro so it lives here and is shared back (Concurro consumes it via a local file: dependency). Hugging Face catalog, the Web Component wrapper, and native probe adapters come next.