ozhyhinas/webnn-catalog

WebNN Catalog: weights and model libraries

Binary artifacts for webnn-catalog, a set of pre-built, pre-tuned browser inference graphs published as data. The catalog lets a web page run real models on WebNN / Core ML and WebGPU without a build step, an npm package, or a native runtime.

The code, recipes and schema live on GitHub. This repository holds only the large binaries they point at — constant blobs for the WebNN graph recipes, and model libraries plus patched runtime bundles for the WebLLM entries. Each entry's entry.json in the GitHub repo references the files here by URL.

Everything was built and measured on an Apple M5 Pro, macOS 26.6.2, Chrome 152.

WebNN graph constants (Core ML)

family	task	derived from	size	latency	fidelity
intellitex-t5-220m	English → LaTeX	duanxianpi/IntelliTex (CodeT5+ 220M)	993 MB	146.2 ms	identical tokens to fp32 on 15/15 items
texify-420	420×420 image → LaTeX	vikp/texify	598 MB	145.7 ms	identical greedy tokens on 18/18 images
texo-384	384×384 image → LaTeX	Texo / FormulaNet (20M)	42 MB	25.9 ms	identical greedy tokens on 18/18 images

For scale, the same Texo model through ONNX Runtime Web takes 770 ms on WASM fp32 and 780 ms on WebGPU. The catalog entry takes 25.9 ms.

WebLLM model libraries (WebGPU)

Compiled WebGPU model libraries and hash-pinned runtime bundles for greedy decoding with device-resident argmax and batched command encoding. Model weights are not duplicated here; entries point at the upstream MLC repositories at a pinned revision.

family	artifacts
qwen3-0.6b-q4f16-1, qwen3-1.7b-q4f16-1	model library + patched WebLLM 0.2.84 runtime bundles
qwen3-4b-q4f16-1, qwen3-8b-q4f16-1	model library
qwen3.5-0.8b-q4f16-1, qwen3.5-4b-q4f16-1, qwen3.5-9b-q4f16-1	model library
minicpm5-2b-q4f16-1	model library

SD-Turbo lives in a separate repository because of its size: ozhyhinas/webnn-catalog-sd-turbo.

Use it

The loaders are two dependency-free ES modules in the GitHub repository. Vendor them or serve them from your own static host.

import { loadEntry, createEntryTensors, assertCoreMLFingerprint } from "./runtime/loader.js";

const dir   = "/families/texo-384/entries/coreml-apple-m5-pro-macos26-chrome152";
const entry = await (await fetch(`${dir}/entry.json`)).json();

const ctx = await navigator.ml.createContext({ deviceType: "gpu" });
assertCoreMLFingerprint(ctx);           // fail loudly, not slowly

const rig = await loadEntry(entry, "https://huggingface.co/datasets/ozhyhinas/webnn-catalog/resolve/main", ctx, { baseUrl: dir });
const t   = await createEntryTensors(ctx, rig);

WebLLM entries use the other loader, which verifies the SHA-256 of the runtime JavaScript and model library before running either:

import { loadWebLLMFromUrl } from "./runtime/webllm-loader.js";

const rig = await loadWebLLMFromUrl(entryUrl);
const out = await rig.engine.chat.completions.create({
  messages: [{ role: "user", content: "Hello" }],
  temperature: 0,
  max_tokens: 64,
});
await rig.dispose();

Full documentation, the recipe schema and every entry's metadata: https://github.com/OlehZhyhinas/webnn-catalog

Notes

The catalog stores and describes; it does not select. There is no ranking, no fallback and no default entry. Probing the machine and choosing an entry is the consuming application's job. An entry's target is what it was built for; compat.requires is what must hold for it to work at all, and those are different claims.

Used in production by LatexGen, which converts spoken math and equation screenshots to LaTeX entirely on-device.

---

Built and published by Oleh Zhyhinas (GitHub, LinkedIn).