LLMBOTTLENECK.COM
For shops and hardware configurators

The question every GPU buyer asks

One script tag tells a visitor which models fit in the memory of the machine on your product page, with an estimated speed — each figure labelled published or estimated, and runtime support stated as untested.

Install it
htmlwidget.js
<script src="https://llmbottleneck.com/widget.js"
  data-hardware="nvidia-geforce-rtx-4090"
  data-context="8192"></script>

That is the whole integration: nothing to keep secret in a public page template.

Try it on your own product

Build the tag
Paste thishtml

<script src="https://llmbottleneck.com/widget.js"
  data-hardware="nvidia-geforce-rtx-4090"
  data-context="8192"></script>
Live, not a mock-upthe same route the tag loads

Why a shop wants this

It answers the objection that stops the sale

“Is 16 GB enough?” is the question that sends a buyer to a forum and loses them. Answering it on the product page, with the source shown, keeps them there.

It sells the bigger card honestly

When a model does not fit, the widget names the ways out cheapest first — a smaller quantization, offload, then more memory. A buyer who was told the truth about the cheap option trusts the expensive one.

It says what kind of answer it is giving

Memory and bandwidth come from the manufacturer’s specification, for the exact configuration the page sells, and architecture from the model’s config.json. The memory verdict and the decode speed are estimates and are labelled as estimates; whether a given runtime loads the model on that machine is not tested, and the result says so. A device or model the widget does not recognise is named as such rather than swapped for another.

What your developer will ask

Does it touch my CSS?No. It renders in an iframe, so your theme cannot reach in and our styles cannot leak out. A theme update can never break it.
Cookies or tracking?None. The widget sets no cookie, uses no storage, and posts one message out of the frame: its own pixel height, so it can size itself to your layout. Nothing about the visitor leaves the frame.
Does it need a key?No, and it does not accept one. A public page template is the wrong place for a secret.
How heavy is it?Under a kilobyte of script, and the frame itself is lazy-loaded — nothing downloads until it scrolls near the viewport.
What if scripting is off?Nothing renders and nothing breaks. Add your own link to https://llmbottleneck.com inside a <noscript> if that matters to you.
Which devices and models?135 devices and 327 models as of the 2026-10-03 snapshot. If you sell something not in the list, tell us — a device with a published specification gets added.
Attributesdata-hardware · data-memory · data-variant · data-usable · data-model · data-quantization · data-context · data-kv · data-batch · data-devices · data-height · data-max-width · data-ref
Free, or under your own brand

Free on any site, shops included. Paste the tag and it works, with the LLM Bottleneck name and a one-line credit under the frame.

Pro — $19 a month. The widget runs without our branding on your domain, and results from the API can be shown without the credit.

Business — $79 a month. Unbranded on up to ten domains, with your own button in the frame: a store search, a product page or an affiliate link, filled in with the model and graphics card the reader just chose.

See the plans →

Prefer to build it yourself?

The widget is a wrapper over the same engine the API exposes. If you want the numbers inside your own layout rather than in a frame, take a free key and call /v1/analyze directly — no permission needed, and the response carries the same evidence tiers and the same refusals.

llmbottleneck
catalogue 2026-10-03models 327devices 135