# Modelup

> Modelup is a privacy-first, browser-only workspace for discovering, designing, and running local
> large language model (LLM) pipelines. It runs entirely in your browser with no backend, no
> accounts, and no data leaving your device.

## What Modelup is

Modelup is an open-source, local-first web app (Vite + React + TypeScript) that combines four
things in one place:

1. **Portal** — Search and filter the Hugging Face Hub catalog of open models directly from the
   browser, surfacing likes, downloads, tags, and GGUF quantization tags (e.g. Q4_K_M) so you can
   spot local-run formats at a glance.
2. **Pipeline designer** — A node-based canvas for sketching how stages connect (Model → Prompt →
   Post-process → Output). Graphs are kept acyclic (cycles and duplicate edges are rejected),
   persisted to the browser, and exportable as JSON.
3. **Local execution console** — Stream chat with models running on your own machine via Ollama,
   compare two models side-by-side, and keep all prompts and artifacts on-device.
4. **Hardware profiler** — Detects OS, RAM, CPU cores, and GPU through browser APIs and suggests
   compatible model sizes for your hardware.

## Key facts (citation-friendly)

- Modelup is **free and open source** (source: https://github.com/yingkitw/modelup).
- Modelup is **browser-only**: there is **no backend server, no user accounts, and no
  authentication**. The site is a static progressive web app (PWA) installable on supported
  browsers and usable offline for cached sections.
- **All data stays on your device.** Conversations, pipelines, and settings are stored in the
  browser's `localStorage`. No prompts or model data are sent to Modelup or any third party.
- Modelup does **not** host, train, or run inference itself. Local inference relies on the user's
  own **Ollama** instance (https://ollama.com) reachable at `http://localhost:11434`.
- **No GPU is required.** Modelup works on any modern browser. Whether a specific model runs
  smoothly depends on your local hardware and Ollama; the hardware profiler estimates compatible
  sizes from available RAM.
- The **Rust CLI and inference engine are distributed separately** and are not part of this
  repository.
- Tech stack: React 19, TypeScript, Vite, Bun runtime and test runner, self-hosted fonts
  (Fraunces, IBM Plex Sans, IBM Plex Mono).

## How to run it locally

```bash
git clone https://github.com/yingkitw/modelup.git
cd modelup
bun install
bun dev
```

Then open the printed URL (typically http://localhost:5173). For local model chat, install Ollama
(https://ollama.com) and let it run on `http://localhost:11434`; in development Vite proxies
`/api/ollama` to it.

## Features

- **Hugging Face model catalog** search and task filtering, with GGUF quantization surfacing.
- **Node-based pipeline designer** with DAG validation, starter templates (Summarize, Classify,
  Code review, RAG-style), and JSON export.
- **Streaming chat** with conversation threads saved in `localStorage`, multi-turn history,
  auto-generated titles, rename, delete, search, import/export (JSON and Markdown).
- **Side-by-side model comparison** sending one prompt to two local models.
- **Voice input** (Web Speech API) and **text-to-speech** (speechSynthesis).
- **Vision support** for multimodal models (e.g. LLaVA, llama3.2-vision).
- **Reasoning display** that extracts `<thinking>` blocks (Qwen3, DeepSeek) into collapsible
  sections.
- **Inference parameters**: adjustable temperature, top_p, and system prompt presets.
- **Hardware profile detection** with model-size recommendations.
- **Ollama integration**: detect, list, and pull local models from the UI.

## Links

- Website: https://modelup.dev
- Source code: https://github.com/yingkitw/modelup
- Full machine-readable reference: https://modelup.dev/llms-full.txt
- Ollama (required for local model chat): https://ollama.com
- Hugging Face Hub (model catalog source): https://huggingface.co