llmfit-web

A web-based version of llmfit — find what LLM models run on your hardware, right in your browser.

🌐 Live Demo: https://rafaelmaza.github.io/llmfit-web/

---

⚠️ Project Origin

This repository was AI-generated as a web port of the original llmfit Rust CLI tool.

All credit for the core algorithm, scoring logic, and conceptual design goes to AlexsJones, the creator of llmfit.

This project reimplements the llmfit scoring engine in JavaScript to make it accessible via a web interface. The original Rust implementation remains the authoritative reference.

---

Overview

llmfit-web brings the model-scoring engine from the original Rust CLI tool to the web, allowing anyone to input their hardware specs and get instant recommendations for which LLM models will run well.

Features

• 50+ GPU models — NVIDIA RTX/A-series, AMD RDNA/MI-series, Intel Arc, Apple Silicon
• 206 LLM models — From 812K to 753B parameters, including MoE architectures
• Multi-dimensional scoring — Quality, Speed, Fit, and Context analysis
• Smart quantization — GGUF and MLX quantization hierarchies with automatic optimization
• Real-time ranking — Instant results, no backend needed
• Mobile responsive — Works on phones, tablets, and desktops

How It Works

1. Select your GPU (or enter custom VRAM)
2. Choose your system RAM (8GB to 128GB)
3. Pick your use case (Coding, Reasoning, Chat, Multimodal, etc.)
4. Get ranked recommendations — Top models scored by fit, speed, and quality

The scoring engine analyzes:

• Memory fit — Does it fit in VRAM, need CPU offload, or use MoE expert offloading?
• Speed — Tokens/sec based on backend (CUDA, ROCm, Metal, CPU) and quantization
• Quality — Parameter count, model family, quantization penalty, task alignment
• Context — Context window availability vs. use-case requirements

Run Modes

• GPU — Model fully loaded into VRAM (fastest)
• MoE Offload — Active experts in VRAM, inactive experts in system RAM
• CPU Offload — Partial GPU, model spills to system RAM
• CPU Only — Entire model in system RAM (slowest)

Fit Levels

• Perfect — Recommended memory met on GPU
• Good — Fits with headroom
• Marginal — Tight fit or CPU-only
• Too Tight — Does not fit

Project Structure

llmfit-web/
├── src/
│   ├── scoring.js      # Core multi-dimensional scoring engine
│   ├── gpus.js         # GPU lookup table (NVIDIA, AMD, Apple, Intel)
│   ├── utils.js        # Quantization hierarchy, speed calculations
│   └── models.js       # Model database (206 models)
├── test/
│   └── scoring.test.js # Unit + integration tests (30 passing)
├── index.html          # Web UI (production-ready)
├── server.js           # Local dev server (Node.js)
├── package.json        # Dependencies
└── README.md           # This file

Running Locally

Start the dev server:

npm start

Then visit:

• Local: http://localhost:8000
• Network: http://:8000

Or open index.html directly in a browser (works without a server).

Testing

Run the test suite:

npm test

Current: 30/30 tests passing ✅

• 23 unit tests (scoring engine, memory fit, speed, quality)
• 7 integration tests (real-world hardware + model combinations)

Deployment

This project is static HTML + JavaScript — no backend, no build step, no runtime dependencies.

Deploy anywhere:

• GitHub Pages (built-in)
• Netlify / Vercel (drag-and-drop)
• Cloudflare Pages
• S3 + CloudFront
• Or just serve the files from any web server

Recording a Demo GIF (optional)

If you want a short GIF for the README, the repo includes a Playwright script that drives a headed Chromium session and records a demo video.

1) Install dev deps

npm install

2) Record (headed)

Local machine with a GUI:

npm run record:demo

Headless Linux server (uses a virtual display):

sudo apt-get install -y xvfb
xvfb-run -a npm run record:demo

The video will be saved under:

artifacts/videos/

3) Generate the GIF (easy)

This repo includes a script that captures PNG frames via Playwright and encodes an animated GIF using gifenc + pngjs:

npm run generate:gif

This writes demo.gif in the repo root.

Model Database

206 open-source models included:

Category	Count	Parameter Range
Tiny	31	<1B
Small	44	1-7B
Medium	66	7-20B
Large	37	20-70B
XLarge	28	70B+
MoE	33	Mixture of Experts

Source: Extracted from HuggingFace via the original llmfit project.

API (for developers)

Use the scoring engine programmatically:

import { scoreModels, UseCase } from './src/scoring.js';
import { getAllModels } from './src/models.js';

const hardware = {
  gpuVramGB: 24,
  systemRamGB: 64,
  backend: 'CUDA'
};

const useCase = UseCase.Coding;

const results = scoreModels(getAllModels(), hardware, useCase);
console.log(results.slice(0, 5)); // Top 5 models

Credits

Original Project: llmfit by AlexsJones

This web port was AI-generated and released under the same license as the original project.

License

MIT License (matching the original llmfit project)

---

Contributing

Issues and PRs welcome! If you find a bug or have a feature request, please open an issue.

For major changes to the scoring algorithm, consider contributing to the original llmfit project first.