New in 3.0: a redesigned browser workspace. Every feature is still there, but each control appears only when you need it, and the Models view now loads the full ranked catalog. See what changed →
Let Sparky check your rig
RigSpark keeps the first question simple: will this model actually run here? The answer comes from a curated offline catalog and your detected hardware, not from a download you have already waited for.
1. Install RigSpark
Homebrew on macOS or Linux:
On a Mac you can also install the desktop app: download RigSpark_aarch64.dmg
(Apple Silicon) or RigSpark_x86_64.dmg (Intel) from the
latest release
and drag RigSpark to Applications. The app is signed with a Developer ID and
notarized by Apple, so it opens without a Gatekeeper warning. The app bundles llama.cpp, so it needs no other
runtime: on first launch, Download & start fetches a verified starter model
that fits your Mac. It lives in the menu bar and launches at login (turn that
off in Tools); closing the window keeps your model serving, and
Quit RigSpark stops it. Updates install after their signature is verified.
On Windows, download RigSpark_x64-setup.exe from the
latest release
and run it. It installs for the current user without admin rights. The installer
is not code-signed yet, so if SmartScreen warns, choose More info → Run
anyway. The Windows app bundles a CPU build of llama.cpp, and its advice for
that runtime counts system memory only.
For the command line on Windows and other platforms, download a
prebuilt archive,
extract it, and add the folder to PATH. Keep rigspark, llmup, and
rigspark-gui together.
The RigSpark binaries need no Node.js, Python, or compiler. Inference backends have their own requirements.
If Sparky helps you, please consider sponsoring the project: https://buymeacoffee.com/shashanksw9
Cargo, checksums, unsigned macOS archives, and upgrades → · Docker caveats →
2. Sparky sizes up your hardware
Start with advice. It is deterministic, works offline, and does not require an inference backend:
recommendranks the catalog for this machine.can-runchecks one model before you download it.catalog --alllets you browse the bundled model data.
Use --json for automation and --accessible for screen-reader-friendly
output. llmup remains a compatibility alias.
3. Read the face: yes, slow, or no
| Verdict | What Sparky means |
|---|---|
| yes | The model fits with the required headroom. |
| slow | It fits, but the expected experience may be constrained. |
| no | It does not safely fit the detected memory budget. |
Every verdict includes the reason. When RigSpark cannot source a figure, it says
unknown rather than inventing one. Throughput ranges are estimates, not
benchmarks.
4. Pick a model, verify it, and start
Once you have a fit, move from advice to a running model:
RigSpark supports:
- Ollama — recommended default and managed child process
- llama.cpp — native local serving
- MLX — Apple Silicon
- LM Studio — attach to an already running server
- BitNet — Microsoft's I2_S weights on CPU, using your compiled bitnet.cpp server
Managed weights pass an integrity check before serving. Servers bind to
127.0.0.1 by default.
For Microsoft's official BitNet b1.58 2B-4T:
Build the server using Microsoft's instructions. On Apple Silicon, use the tested pinned build and ARM patch; unmodified builds can produce invalid output or crash, and BLAS must be disabled. Stock llama.cpp cannot load this I2_S artifact. RigSpark downloads and verifies the weights, not the runtime. Keep the variable set for chat, doctor and shutdown. The initial integration is CPU-only; speed and KV-memory estimates stay unknown. See the per-model benchmark reports for the recorded BitNet chat results, methodology, and limitations.
Backend setup and lifecycle details →
Faster serving: speculative decoding and parallel agents
--speculative auto|draft|ngram|off. On llama.cpp, models with a catalog drafter use it by default. RigSpark downloads the drafter, verifies its SHA-256 and reserves its memory before checking fit. If the drafter does not fit, RigSpark starts without it.ngramdrafts from the context and needs no extra weights. Ollama handles speculative decoding inside its own model tags.--parallel N(1–16) sets concurrent request slots. llama.cpp slots share one unified KV pool. Ollama allocates a full context for each slot, so it needs--contextand known KV geometry to size the memory (or--bypass).rigspark benchmeasures the active server with fixed, versioned prompts: time to first token, prefill and decode tok/s, draft acceptance, and combined throughput across concurrent requests. The prefill, decode and draft figures come from the runtime's own counters; anything it does not report shows asunknown.
Connect your coding agent
connect points the agent at the active loopback server for that launch only.
It never rewrites the agent's own config files. Add --dry-run to print the
command and environment instead of starting the agent.
5. Continue in the browser workspace
Choose a model, inspect the same hardware verdicts, chat, and use agents, skills, and MCP tools from the loopback-only browser workspace.
- One model chip in the top bar holds the model, runtime, harness, and agent pickers.
- Models lists the whole ranked catalog with search, Runs well / Slow / Won't fit filters, and a Fit settings popover for context and KV cache.
- The inspector (⌘I, or click Sparky) shows live memory, CPU, disk, and the latest call's token usage. Sparky's mood follows real RAM pressure.
- Light and dark themes, with ⌘1–⌘5 to switch views.
Local chat stays on your machine. Cloud harnesses and external tools can send data to their own providers, so RigSpark keeps those boundaries visible.
Browser workspace, agents, and tools →
6. Let Sparky create images and short videos
RigSpark can orchestrate open-weight image and video models through your own ComfyUI installation.
Start ComfyUI on its default loopback address, 127.0.0.1:8188. RigSpark
downloads pinned weights, verifies their SHA-256 digest, installs only its
built-in local workflows, and refuses paid Partner API nodes. Generation speed
stays unknown when no defensible measurement is available.
In rigspark gui, switch the composer from Text to Image or Video.
Generated media renders inline and is never sent to the language model.
For UI/UX mockups, select ming-image:ming_image_0.1_design_int8_convrot under Models → Image → Use in Chat, then generate in Image mode. This exact 25.94 GB weight pairing was tested with ComfyUI 0.38.0 on a 36 GiB Apple Silicon Mac. Other Ming variants remain fit-only; NVIDIA/CUDA and peak inference memory are not certified. Generated lettering can be garbled.
In the GUI, switch the composer to Image and pick a shape (Thumbnail
16:9, Wallpaper 9:16 or Square) and how many versions (1, 2 or 4),
each on its own seed. Keep a seed to steer one picture with new words. Then
download it, copy it into any app, or Send to app…: with Canva, Figma or
another connector added under Connectors, the chat hands the file to that
app's tools (each call asks for your approval) for text, layouts and resizing.
RigSpark serves the model; editing stays in the tools you already use.
Pictures are made one at a time, and the time per picture is measured on your
machine after the first one. From the CLI, pass --aspect landscape or
--aspect portrait to rigspark generate image.
Image and video generation specification →
7. Talk to it: spoken replies and dictation
In rigspark gui, open Models → Voice and start a voice model:
- Speak (text to speech):
kokoro-82morchatterboxthrough mlx-audio on Apple Silicon (pip install -U mlx-audio). - Listen (speech to text): Whisper large-v3-turbo, small or base.en through
whisper.cpp's
whisper-server(brew install whisper-cpp, or setRIGSPARK_WHISPER_SERVER). RigSpark downloads the pinned ggml file and verifies its SHA-256 before loading it.
Then turn on Speak replies in the composer, press the microphone to
dictate, or switch the composer to Voice to have a script written and
read aloud, or to read your exact text. Audio plays inline and downloads as
WAV. On Linux or Windows, attach any loopback server with OpenAI's
/v1/audio/speech or /v1/audio/transcriptions (Kokoro-FastAPI, LocalAI)
under Runtime → Voice settings. Voice servers bind to 127.0.0.1, fit
covers weight memory only, and voice speed stays unknown until measured.
Why the answers are trustworthy
- Offline advice. Recommendation commands use a bundled, curated catalog and make no network calls.
- Honest unknowns. Missing bandwidth, model geometry, or measurements stay
visibly
unknown. - Fail-closed integrity. Managed weights must match their catalog digest, or a documented size-floor fallback, before RigSpark serves them.
- Loopback by default. Local servers listen on
127.0.0.1, not the public network. - Reproducible decisions. The same hardware and catalog produce the same recommendation.
- One adapter boundary. Ollama, llama.cpp, MLX, LM Studio, and BitNet sit behind the same runtime lifecycle and safety gates.
Add any GGUF model from Hugging Face
The bundled catalog is curated and signed. For anything newer or more specific, add it yourself:
In the browser workspace, use Models → Add from Hugging Face.
RigSpark reads the repository and one GGUF header to size every
quantization against your machine. It then pins the repository commit and
each file's SHA-256 in ~/.rigspark/catalog/local.json. up downloads
through llama.cpp or Ollama (hf.co/owner/repo:quant) and refuses to
serve weights whose digest does not match. Repositories need a recognized
open-weight license. Projector files and split (multi-part) GGUFs are
skipped.
Added models appear in recommend, can-run and the GUI marked added by
you. Remove one with rigspark add --remove <id>. Network access happens
only when you search or add; advice stays offline.
Update the catalog when you choose
Catalog updates are explicit; recommendations and normal startup remain offline.
RigSpark verifies the signed catalog before activating it atomically. An invalid signature, incompatible schema, or network failure leaves the current snapshot unchanged. Recovery falls back through the previous verified snapshot and then the bundled catalog, with visible warnings.
New models arrive weekly. A scheduled pipeline admits every new Ollama library
model whose facts it can read from the shipped artifact (size, digest, context,
architecture, open license), plus Comfy-Org image and video releases, and
publishes a signed catalog once the quality gate passes. These entries are marked
auto-sourced, keep fail-closed digest verification, and show unknown for
anything that could not be sourced. Image and video models without a built-in
workflow yet get a fit verdict and are labelled workflow coming.
In the terminal UI, m cycles the same 1, 2 and 3 month windows; the browser
workspace has a Released in selector.
rigspark catalog --refresh is a maintainer preview of enrichment against the
bundled registry snapshot; it does not install a published catalog.
Catalog provenance, trust, and recovery →
A 20-second tour
The tour runs rigspark recommend, shows models that do and do not fit, then
opens the same verdicts and a local chat in rigspark gui.
Watch in 1080p with sound · Earlier demo on YouTube
Find your next command
| Goal | Command |
|---|---|
| Rank models for this machine | rigspark recommend |
| Check one model | rigspark can-run <model> |
| Diagnose hardware and runtimes | rigspark doctor |
| Browse text models | rigspark catalog --all |
| Browse image and video models | rigspark catalog --generation |
| Start a verified model | rigspark up <model> |
| Open the terminal UI | rigspark |
| Open the browser workspace | rigspark gui |
| Generate local media | rigspark generate |
| Benchmark the active server | rigspark bench |
| Connect a coding agent | rigspark connect <agent> |
| Add a model from Hugging Face | rigspark add <owner/repo> |
| Stop the active model | rigspark down |
Run rigspark --help for the complete CLI.
Project map
crates/rigspark-core/— catalog, memory sizing, ranking, advice, and reportscrates/rigspark-runtime/— hardware detection, backend adapters, lifecycle, state, and harnessescrates/rigspark-gui/— loopback HTTP/SSE host and embedded browser clientcrates/rigspark-cli/— native CLI, terminal UI, and catalog maintenanceapps/desktop/src-tauri/— native desktop shell
Build from source
RigSpark is a Rust workspace pinned to Rust 1.98.1:
No Node.js or TypeScript tooling is part of the project.
Want to help? Read CONTRIBUTING.md and pick an issue labelled
good first issue.
Keep exploring
- Commands, custom context, and installed models
- Terminal UI
- How hardware advice works
- Scripting and exit codes
- FAQ and troubleshooting
- Development guide
- Project specification
- Changelog
Support RigSpark
RigSpark is free, MIT-licensed, and built in spare time. If it saved you a download that would never have run, you can buy me a coffee ☕.
Keywords: local LLM, run LLM locally, LLM hardware requirements, VRAM calculator, tokens per second, Ollama, llama.cpp, MLX, LM Studio, Rust CLI.
