What is Thumper-Run?

Thumper-Run is a local-first AI app launcher. It runs image generators, language models, music tools, video pipelines, and AI agents on your hardware with one-click install, automatic GPU detection, and opt-in network features.

Local work stays on your device until you choose a sync or cloud workflow. Current relay-assisted workflows are server-assisted and should not be treated as private end-to-end encryption.

No account required. No API keys. No usage limits. Install an app, download the models, and generate — offline works too.

5 Minutes to Thumper-Run

bash

# 1. Download the release installer for your OS
# Open https://thumper.run/download and verify the published checksum when available
# ✓ Install using the native macOS, Windows, or Linux package
 
# 2. Launch Thumper-Run
thumper-run
# ✓ GPU detected: NVIDIA RTX 4070 (12 GB VRAM)
 
# 3. Install an app from the catalog
# Click: Catalog → ComfyUI → Install
# ✓ Cloned, pip installed, models downloaded
 
# 4. Generate your first image
# Click: Launch → Load default workflow → Queue Prompt
# ✓ Image saved to Gallery

That's it. No Python setup, no pip install, no model hunting.

Manual vs Thumper-Run

Step	Manual Install	Thumper-Run	Time Saved
Install Python	Download, PATH, venv	Automatic (bundled)	~15 min
Install app	git clone, pip install	One click	~10 min
Download models	Find URLs, wget, verify	Auto-downloaded	~20 min
Place models	Read docs, move files	Symlinked automatically	~5 min
Configure GPU	CUDA/ROCm flags, env vars	Auto-detected	~10 min
Start the app	Terminal, correct args	One click	~2 min
Total	30–90 minutes	~5 minutes	25–85 min

What Thumper-Run automates: Python environment isolation, GPU backend selection, model downloading and placement, dependency resolution, health checks, and process management.

What stays the same: you still choose which apps to run, which models to use, and how to configure your workflows. Thumper-Run handles the plumbing, not the creativity.

Key Concepts

Catalog

A curated directory of AI apps packaged for one-click install. Each entry includes a manifest, model packs, and platform patches. See App Manifests →

Manifest

A .thumper.yaml file that describes how to install, configure, and launch an app. It declares the runtime, dependencies, GPU patches, and model requirements. Tutorial: App Manifest →

Model Pack

A bundle of AI model files (weights, configs, VAEs) described by a .pack.yaml manifest. Packs handle downloading, SHA-256 verification, and VRAM-aware variant selection. Tutorial: Model Pack →

CRDT Sync

Conflict-free replicated data types power multi-device sync. App settings, gallery metadata, and preferences merge automatically without conflicts — even after days offline. Platform: CRDT Sync →

GPU Detection

On first launch, the platform probes your hardware (NVIDIA, AMD, Intel, Apple Silicon, CPU) and configures PyTorch indexes, environment variables, and runtime flags automatically. Installation: GPU Auto-Detection →

Patches

Platform-specific config overrides applied at install or launch time. Patches adjust CLI args, environment variables, and config files for different GPUs and operating systems. App Dev: Platform Patches →

Health Check

After launching an app, Thumper-Run polls its HTTP endpoint until it responds. The UI shows a spinner until the health check passes, then opens the app in a webview or browser. Users: Running Apps →

Gallery

A cross-app image gallery that auto-discovers generated images from all installed apps. Extracts ComfyUI, A1111, and Fooocus metadata for searchable generation parameters. Users: Image Gallery →

Hardware Guide

Use Case	Min GPU	Recommended	Example Apps
Image (SDXL)	6 GB VRAM	8 GB+ VRAM	ComfyUI, Fooocus
Image (FLUX)	12 GB VRAM	16 GB+ VRAM	ComfyUI
LLM 7B	4 GB VRAM or CPU	8 GB VRAM	SillyTavern, Ollama
LLM 70B	24 GB VRAM	48 GB+ VRAM	SillyTavern, Ollama
Music	8 GB VRAM	12 GB+ VRAM	ACE-Step
Video	12 GB VRAM	24 GB+ VRAM	ComfyUI (LTX-Video)
CPU-only	None	16 GB RAM	Ollama (GGUF), OpenVoice

GPU Compatibility Matrix

GPU	Linux	macOS	Windows
NVIDIA RTX 30xx/40xx	Yes	N/A	Yes
NVIDIA GTX 10xx/20xx	Yes	N/A	Yes
AMD RX 7000 (RDNA 3)	Yes	N/A	Limited (ROCm)
AMD RX 6000 (RDNA 2)	Partial	N/A	Partial
AMD Ryzen AI (iGPU)	BF16 only	N/A	No
Intel Arc A770/A750	Partial	N/A	Partial
Apple M1/M2/M3/M4	N/A	Yes	N/A
CPU only	Yes	Yes	Yes

AMD APU (gfx1150): FP16 causes NaN in DiT attention. Thumper-Run forces BF16 automatically, but custom workflows must not override the dtype. See Troubleshooting for details.

What It's NOT

Not a cloud API — everything runs on your machine. There are no per-request charges, rate limits, or data retention policies.
Not a model trainer — Thumper-Run launches inference apps. For fine-tuning or LoRA training, use the underlying app directly.
Not a container runtime — apps run natively (Python venvs, Node, or binaries). Docker is not required, though Docker-based apps are supported via manifests.
Not a model hub — models are downloaded from HuggingFace, CivitAI, or direct URLs as declared in model packs. Thumper-Run manages placement and linking, not hosting.
Not an IDE or notebook — there is no built-in code editor or Jupyter integration. Use your preferred editor alongside Thumper-Run.

Learn More

Getting Started

Install Thumper-Run and launch your first app

Tutorials

Step-by-step guides for images, chatbots, music, and more

Platform Architecture

CRDT sync, encryption, offline mode, and cloud compute