What you can do
- Chat with open models such as Llama, Qwen, Gemma and DeepSeek, fully offline.
- Make images and short videos from a prompt, in the same chat window.
- Talk to it: have replies read aloud, or dictate instead of typing.
- Know before you download. Every model gets a yes / slow / no verdict for your RAM and GPU, with the reason, so you don't wait for a 20 GB download that won't run.
- Add any model from Hugging Face. Paste a repository and RigSpark sizes each version against your machine, then picks the largest one that fits.
- Hand off to your tools. Copy a picture into any app, or send it to Canva, Figma and other connectors you add.
No GPU required: small models run on the CPU. On a Mac, RigSpark uses the Apple Silicon GPU; on Linux, an NVIDIA GPU through Ollama or llama.cpp. Each model shows its download size and memory need before you start.
Get started
- Download and open the app. The Mac app is signed and notarized by Apple. The Windows installer needs no admin rights; it isn't code-signed yet, so if SmartScreen warns, choose More info → Run anyway.
- Click Download & start. RigSpark picks a starter model that fits your computer, verifies the file, and starts it. Nothing else to install for chat.
- Chat, or press + to switch to Image, Video or Voice. The first time you make a picture, RigSpark offers to set up its image engine for you.
Advice and chat work offline; RigSpark goes online only to download models and updates you ask for. On a Mac the app lives in the menu bar, and updates install when you approve them, after their signature is checked.
Where each feature runs
| Mac (Apple Silicon) | Mac (Intel) | Windows | Linux | |
|---|---|---|---|---|
| Desktop app | ✓ | ✓ | ✓ | browser app (rigspark gui) |
| Chat | ✓ | ✓ | ✓ (CPU) | ✓ |
| Fit check, Hugging Face models | ✓ | ✓ | ✓ | ✓ |
| Images and video, set up for you | ✓ | ✓ | your own ComfyUI | ✓ |
| Spoken replies | ✓ | other server | other server | other server |
| Dictation | whisper.cpp | whisper.cpp | other server | whisper.cpp |
"Other server" means any local speech server with OpenAI's audio API (such as Kokoro-FastAPI or LocalAI), added under Runtime → Voice settings. macOS 12 or newer. How much a model needs depends on the model: the fit check shows it for each one rather than a single minimum.
Linux and the command line
There's no Linux desktop app yet. On Linux, install the command line and run
rigspark gui: it opens the same app in your browser. With Homebrew on macOS
or Linux:
Prebuilt archives for Linux, macOS and Windows are on the
releases page.
Cargo, checksums and upgrades → ·
Docker → ·
llmup remains a compatibility alias.
RigSpark runs models through Ollama, llama.cpp, MLX, LM Studio or BitNet, and
can point Claude Code, Codex or OpenCode at your local model with
rigspark connect. The guide covers every command.
Why you can trust the answers
- Honest numbers. When RigSpark can't source a figure, it shows
unknowninstead of guessing. Speeds are estimates, not benchmarks. - Verified downloads. Model files must match a pinned SHA-256 digest before they run.
- Local only. RigSpark's servers listen on
127.0.0.1, not your network, and it sends no usage data. - Catalog updates when you choose. The model list is signed and only changes when you update it.
Updates only add or change models; recommendations and normal startup remain offline.
rigspark catalog --refresh is a maintainer preview and doesn't install
anything. How catalog updates work →
Community
RigSpark Community is free, MIT-licensed and built in the open. Bug reports, model requests and pull requests are welcome:
- Report a bug · Suggest a feature · Request a model
- Contributing guide and good first issues
- Changelog · Terms of Use · MIT License
If RigSpark saved you a download that would never have run, you can buy me a coffee ☕.
