OnDesk

What fits on the desk. Not in the cloud.

OnDesk helps you choose local AI hardware with fewer guesses. Pick a GPU or Apple Silicon configuration to see which models fit, or paste a Hugging Face repository to find the hardware that can run it.

Every result shows the memory estimate, best available quantization, fit band, and estimated generation speed. The calculations are deterministic and visible instead of being delegated to an AI model.

Try OnDesk

Use the web app

Open the live demo and test OnDesk immediately—no installation or account required. Choose your hardware from the catalog, then explore compatible models or search in the opposite direction from model to hardware.

Run it on your device

Running OnDesk locally enables host hardware detection in addition to the full web experience. Requirements: Node.js 20 or newer.

git clone https://github.com/hex41434/OnDesk.git
cd OnDesk
npm install
npm test
npm run dev

Open http://localhost:3000.

Other useful commands:

npm run typecheck
npm run build

What OnDesk does

• Hardware to models: choose a device and get a ranked list of compatible language, vision, detection, segmentation, speech, and embedding models.
• Model to hardware: paste an org/model identifier or Hugging Face URL and see the smallest compatible hardware in the catalog.
• Inference and training modes: compare normal inference with a conservative full-training memory estimate.
• Transparent memory math: inspect weights, KV cache, runtime overhead, extra modality memory, and usable device memory.
• Practical filters: narrow results by task, model name, hardware name, memory use, fit quality, quantization, or estimated speed.
• Optional smart search: Gemini improves the search experience by translating informal or multilingual queries into hardware and Hugging Face search terms. It is not required and never decides the fit score.

Deploy to Cloudflare Workers

The full-stack app deploys through OpenNext. A custom domain is optional; Cloudflare provides a *.workers.dev URL automatically.

npm run build:cf
npm run deploy

Host hardware detection remains local-only. Public visitors choose their hardware from the catalog instead of reading the Cloudflare server runtime.

Optional Gemini search

OnDesk works without Gemini: local catalog matching and direct Hugging Face Hub search remain available. Gemini is an optional search-quality enhancement for natural-language queries. To enable it, either save a Gemini API key on the Settings page or create apps/web/.env.local:

GEMINI_API_KEY=your_key_here

A key entered through Settings is stored in that browser's local storage and is sent to the OnDesk API only when smart search is used. Fit calculations remain inside @ondesk/core and do not use Gemini.

How scoring works

At a high level, OnDesk estimates:

working set = model weights + KV cache + runtime overhead + modality extras

The scorer then compares the working set with usable VRAM or unified memory, selects the highest-quality supported quantization that fits, and assigns one of four bands:

• perfect — uses at most 50% of available memory
• good — uses at most 75%
• tight — uses at most 90%
• no — does not fit under the current assumptions

Estimated tokens per second use memory bandwidth and a hardware-specific efficiency factor. They are always labeled est. because they are projections, not benchmark results. The full formulas and assumptions are available at /about while the app is running.

Project structure

apps/web/              Next.js application and API routes
packages/core/         deterministic fit and ranking engine
packages/core/test/    scoring and catalog tests
data/gpus.json         curated hardware profiles
data/models.json       curated model catalog
data/digest.json       digest source data

The main API is intentionally small:

fit(model, hardware, options)
// => { band, gb, quant, note, tokensPerSec, breakdown }

Current limitations

• Hardware and catalog specifications are curated and may lag new releases.
• Speed values are estimates, not measurements from the selected device.
• Training mode models full training at roughly 6× FP16 weights; LoRA and QLoRA are not modeled yet.
• Remote Hugging Face repositories may not expose enough metadata for a reliable estimate.
• Host hardware detection is intended for local development and is not used to identify visitors' devices on a public deployment.

License

MIT