RigSpark: Check Which Local LLMs Your Computer Can Run
Which local LLMs can your computer run? Find out before downloading the weights.
RigSpark scores your hardware and gives every model a yes / slow / no verdict,
a memory-fit explanation, and an estimated tok/s range. Then it verifies, serves,
and chats with the model you pick. It also generates images and short videos
locally through your own ComfyUI. It is an open-source, privacy-first Rust CLI
with a terminal UI and a browser workspace for macOS, Linux, and Windows.
20-second tour: rigspark recommend ranks the offline catalog on an arm64 Mac with
34 GiB of usable RAM and lists the models that won't fit; rigspark gui then shows
the same verdicts and a local chat.
Watch in 1080p with sound · Earlier demo on YouTube
Why RigSpark
- Know before you download. Estimates RAM and GPU/VRAM fit for models such as Llama, Qwen, Mistral, and Gemma.
- Honest numbers. Advice uses a bundled offline catalog. Unknown figures stay
unknown; estimates are not benchmarks. - Safe by default. Managed downloads are integrity-checked, and servers bind
to
127.0.0.1. - Bring your backend. Works with Ollama, llama.cpp, MLX (Apple Silicon), and LM Studio (attach-only).
- Images and video, locally. FLUX.1 schnell and Wan 2.1 through your own ComfyUI, from chat, the CLI, or the terminal UI. No cloud nodes or paid credits.
Install
Homebrew (macOS and Linux):
Windows and other platforms: download a
prebuilt archive,
extract it, and add the folder to PATH. Keep rigspark, llmup, and
rigspark-gui together. The binaries need no Node.js, Python, or compiler;
inference backends have their own requirements.
More: Cargo, checksums, unsigned macOS archives, and upgrades · Docker caveats
Quick Start
Advice works without a backend. To serve and chat, install
Ollama or another supported backend.
Use rigspark --help for all commands, --json for scripting, and
--accessible for screen readers. llmup remains a compatibility alias.
Catalog Updates
RigSpark can activate a signed model catalog independently of an application release. Updating is always explicit; recommendations and normal startup remain offline.
The Models view exposes the same provenance and an Update catalog action. Invalid signatures, incompatible catalogs, and network failures leave the current snapshot unchanged. RigSpark can recover through the previous verified snapshot and then its bundled catalog, with visible warnings.
rigspark catalog --refresh is different: it previews maintainer enrichment from
the bundled registry snapshot and does not install a published catalog. New model
admission still requires reviewed evidence; a newer catalog date is not a promise
of complete upstream coverage. See catalog behavior, trust, and launch
availability.
Image and Video Generation
RigSpark can also run open-weight image and video models locally and for free
through your own ComfyUI install.
RigSpark doesn't install ComfyUI; start it on its default loopback address
(127.0.0.1:8188) and RigSpark handles the weights and workflows.
In rigspark gui, switch the chat composer from Text to Image or
Video and describe what you want, for example "create mount fuji". The
result renders inline and is never sent to the language model. A Create view
offers the same with a seed field and an explicit ComfyUI directory.
From the terminal:
| Model | Kind | Output | License | Weights |
|---|---|---|---|---|
flux1-schnell:fp8 |
image | 1024×1024 PNG | Apache-2.0 | 16.1 GiB |
wan2.1-t2v:1.3b |
video | 832×480, 33 frames at 16 fps (~2 s) animated WebP | Apache-2.0 | 9.2 GiB |
- Verified weights. Pinned to a Hugging Face commit, checked against their
SHA-256 on every run, and placed under
ComfyUI/models/<folder>/rigspark/. Interrupted downloads resume and are still fully hash-verified. - Local only. RigSpark submits only its own built-in workflows of core local nodes, refuses ComfyUI's paid Partner (API) nodes, never sends account keys, and connects over loopback only.
- Honest numbers. The fit verdict covers weight memory only, and generation
speed is reported as
unknown. Video is slow on laptops: expect minutes per clip, not seconds. - Apple Silicon. RigSpark uses the euler sampler for Wan on Apple GPUs (the default sampler produced corrupt video there) and unloads the previous model when you switch between image and video.
Outputs are never overwritten. --comfyui-dir defaults to RIGSPARK_COMFYUI_DIR,
then ~/ComfyUI. Details: spec.
Browser Workspace
Run rigspark gui to pick a model that fits and chat with it, with agents,
skills, and MCP tools. Local chat stays on your machine; cloud harnesses and
external tools can send data to their providers.

Star History
References
- Commands, installed Ollama models, and custom context
- Terminal UI · Browser workspace, agents, and tools
- Catalog and maintenance · How advice works
- Scripting and exit codes · Performance measurements
- FAQ · Troubleshooting · Tool comparisons
- Development and testing · Specification · Changelog
Keywords: local LLM, run LLM locally, LLM hardware requirements, VRAM calculator, tokens per second, Ollama, llama.cpp, MLX, LM Studio, Rust CLI.
