"Can it run?" widget

Embed our GPU-fit checker on your site. Visitors pick a local AI model and a GPU; the widget shows the verdict, the VRAM math, and (where available) measured tokens/s. Free, no tracking cookies, one script tag.

Install

Paste this where the widget should appear:

<div id="caih-canitrun"></div>
<script async src="https://compareaihardware.com/widget/can-it-run.v1.js"
        data-caih-container="caih-canitrun"
        data-caih-model="llama-3-1-8b"
        data-caih-gpu="nvidia-geforce-rtx-3090"></script>

Optional attributes: data-caih-theme (light|dark|auto, default auto) · data-caih-ref (your label, [a-z0-9-], aggregate stats only). The version in the filename (.v1.) is permanent — breaking changes ship as a new file, never under v1.

How it works

The snippet renders inside a shadow DOM (style isolation), embeds its model/GPU lists at build time, and fetches verdicts from GET /api/widget/fit?model={slug}&gpu={slug}&quant=q4_k_m&context=4096. Verdicts are cached per visitor for 10 minutes and hard-cached at the edge for a day. Quant and context are fixed in v1.

VerdictMeaning
Fits comfortablyVRAM ≥ requirement × 1.25
Tight fitVRAM ≥ requirement, little headroom
Needs offload or multi-GPUTwo cards (×0.90 efficiency) would cover it
Won't fitRequirement exceeds what this card can address

Terms & attribution

The widget is free to embed with the attribution link intact. Every card carries "GPU fit data by CompareAIHardware" pointing at our model-to-GPU converter; the widget refuses to display results if that link is removed or hidden. Dataset and fit math are additionally licensed under CC BY 4.0.

  • No cookies, no fingerprinting, no third-party requests — the widget talks only to compareaihardware.com.
  • Rate limit: 60 requests/min/IP; the widget backs off automatically on HTTP 429.
  • Responses are public read-only data (no credentials); CORS is open for embedding.
  • We may block abusive origins; normal editorial use will never hit that.

Changelog

v1 (2026-08) — initial release: model + GPU selects, four-state verdict, VRAM math, measured tok/s when available, attribution enforcement.