# browser-llm-fit

> Auto-detect client hardware capabilities and match against browser-executable AI models across WebGPU, WASM, and ONNX Runtime.

## Links
- [Application](https://h3manth.com/ai/browser-llm-fit/): Live interactive hardware matrix and profiler
- [Full Documentation](https://h3manth.com/ai/browser-llm-fit/llms-full.txt): Comprehensive model specifications and agent discovery guide
- [Source Code](https://github.com/hemanth/browser-llm-fit): GitHub repository and automated model registry updater
- [npm Package](https://www.npmjs.com/package/browser-llm-fit): Node / browser utility module
- [Author](https://h3manth.com): Hemanth HM

## Overview
browser-llm-fit probes client browser capabilities including WebGPU support, shader-f16 floating point support, maximum storage buffer binding size, navigator.deviceMemory, and WASM SIMD/threads. It evaluates 30+ machine learning models to determine which can run client-side inside a browser tab with zero server roundtrips.

## Features
- Hardware Auto-Detection: Probes navigator.gpu for adapter limits, shader-f16 half-precision, buffer limits, and CPU concurrency.
- Compatibility Matrix: Evaluates 30+ models across LLM, Vision, Audio, Embeddings, and NLP modalities.
- Hardware Simulation Sandbox: Simulates Apple M3 Max, discrete RTX 4080, Chromebook, and mobile phone environments.
- GPU & CPU Stress Benchmark: Real-time WGSL compute shader matrix multiplication and buffer allocation benchmarking.
- In-Browser Live Inference: Zero-server model execution via Transformers.js v3 ONNX Runtime Web.
- Code Snippet Generator: Ready-to-use implementation code for WebLLM, Transformers.js, Wllama, and TensorFlow.js.

## Supported Runtimes
- [WebLLM](https://webllm.mlc.ai): MLC AI WebGPU WGSL compute shaders (Llama 3.2, Qwen 2.5 Coder, DeepSeek-R1 Distill, Mistral 7B).
- [Transformers.js v3](https://huggingface.co/docs/transformers.js): ONNX Runtime Web with WebGPU and WASM SIMD (SmolLM2, Whisper, Florence-2, BGE-M3).
- [Wllama](https://github.com/ngxson/wllama): llama.cpp compiled to WebAssembly with multi-threading (TinyLlama, Danube 3 GGUF).
- [Google MediaPipe](https://ai.google.dev/edge/mediapipe/solutions/guide): Google on-device Web ML (Gemma 2B, Face Mesh).
- [TensorFlow.js](https://www.tensorflow.org/js): WebGL and WebGPU neural networks (Universal Sentence Encoder, MobileNetV4).
- [Chrome Built-in AI](https://developer.chrome.com/docs/ai/built-in): Gemini Nano via window.ai.languageModel.