Your LLM router or classifier, distilled to run on plain CPU.

Stop paying an LLM to pick from a list.

Send us examples of the decisions your LLM makes. Get back a small model that makes the same calls on any CPU, for a fraction of the cost. It's yours to keep.

Free report in about a minute · No account · $99 only if you keep the model

  1. Upload examples

    Past questions and the answer your LLM gave. One file, no account.

  2. See your savings

    In about a minute: how much of your traffic a CPU model can take, and what that saves you.

  3. Keep the model

    $99, once. Runs on your servers, even offline. Anything it isn't sure about goes to your LLM.

What's inside, and how we measure it →

Using Claude Code? It does the whole thing and asks before uploading.

It finds the decision in your code and extracts past answers on your machine. Add the plugin, then ask it for a NoGPU report.

claude plugin marketplace add doronp/nogpu-mcp
claude plugin install nogpu@nogpu

Read the source (GitHub)

One JSON object per line with the input (text, structured fields, or both) and the answer, for example {"text": "refund my last order", "label": "refunds"}. A JSON array or a CSV with text and label columns works too. 160 to 10,000 rows, at least 2 different answers, up to 20 MB. Converting LLM logs? Use our free script.

No data at hand?

Try a CPU model trained on 3,000 real banking questions: type one in.

Try it

Free audit

$0
  • Up to 10,000 rows per upload
  • What share the CPU model answers alone and how often it matches your LLM, with a 95%-confidence lower bound
  • Savings calculator at your price per call
  • 3 reports a day, no account
Get my free report

Continuous factory

$299/mo

Coming soon

  • Unlimited models
  • Retrained and re-checked as your traffic changes
  • New versions delivered to your registry

One email when it opens. No marketing, no sharing with third parties.

What kinds of decisions work?

Routing, intent and category labels, allow/deny checks and triage: anything with a fixed list of 2 to about 100 answers. If yours isn't a fit, the free report says so.

What happens to my data?

It trains only your model and is deleted 24 hours after upload (24 hours after purchase, so you have time to download). "Delete my data now" on your report removes it at once.

Personal data in my tickets?

Rows are used only to train and test your model, stored in the US (Google Cloud, us-central1) and deleted 24 hours after upload (after purchase if you buy). Strip emails and names first if you prefer; the report works without them.

Am I locked in?

No. You download the model and a small script that runs it, using open-source libraries only. It runs offline, forever.

New category?

Upload again for a new model ($99). Each model learns one fixed list of answers.

What does it need to run?

An ordinary CPU. No GPU, and a few milliseconds per decision.

Why not TRACER or a fine-tuned BERT?

Both are good if you have the time: TRACER (free, open source) needs Python code and an embedding model you run next to it; a fine-tuned BERT needs training and serving. NoGPU is a file upload: in a minute you know, on your own data, what share a CPU model takes alone and a 95% lower bound on how often it matches your LLM, before you write any code or pay anything.

Technical details: the benchmark, the download, the log converter →

Find out in a minute how much of your LLM bill a CPU can take.

Get my free report