browser-llm-fit

Check if an AI model fits the browser hardware (WebGPU, WASM, RAM, shader-f16).

https://github.com/user-attachments/assets/f4ceae3a-a2b3-4955-bec2-36de9238f56a

npm install browser-llm-fit

Quick start

import fit from 'browser-llm-fit';

const res = await fit('SmolLM2-135M');
console.log(res.fits);   // true
console.log(res.tier);   // 'smooth'
console.log(res.speed);  // '45-65 tokens/sec'

fit('model') checks runtime features, buffer limits, and memory requirements, with warnings where capacity is unknown. fit() returns all models sorted by device compatibility. That's the whole API.

Probe all models

import fit from 'browser-llm-fit';

const { hardware, models } = await fit();
console.log(hardware.webgpuDevice);
console.log(models[0].model.name, models[0].evaluation.score);

Returns detected adapter limits and catalog models ranked by compatibility score.

Simulate constraints

import fit from 'browser-llm-fit';

const res = await fit('Llama-3.2-3B', { ram: 4, gpu: 'wasm-cpu' });
console.log(res.fits); // false

Simulates low-memory phones, Chromebooks, or WASM CPU fallback without physical hardware.

Demo

• Live: https://h3manth.com/ai/browser-llm-fit/

npm run dev

Starts the local development server on http://localhost:5180.

License

MIT © Hemanth.HM

Compatibility limits

Results are estimates, not a guarantee that model allocation or inference will succeed. Browsers do not expose free VRAM. GPU models therefore receive a memory warning unless you supply an explicit budget, for example:

await fit('Qwen 2.5 7B', { ram: 32, vram: 8, gpu: 'webgpu-f16' });

RAM may also be estimated when the browser does not report it. Throughput ranges are heuristics, not measurements of your device. cpuCores is retained as simulation metadata; it does not calibrate throughput. Empty options preserve automatic detection. The API requires a browser environment, including when simulating constraints.

Live demos support the text models listed in the demo selector. Audio models have code examples but no embedded demo until an audio input flow is implemented. Inference runs in a worker; switching models or closing the dialog terminates it. Completed runs dispose the pipeline, and subsequent runs load it again using browser caches.

Development checks

npm run lint
npm test
npm run build:app
npx playwright install chromium
npm run test:browser

The browser check starts and stops its own preview server. Set TEST_BASE_URL to check an already running server instead. The weekly registry job writes discovery candidates to src/data/registry-sync.json; catalog additions still require review.