New RigSpark 3.0.6: now on Windows

Meet Sparky.

Know which local LLMs
your computer can run.

Sparky reads your RAM, GPU and disk, then gives every local model a yes, slow or no with estimated tok/s. Advice runs fully offline.

Download for macOSCommunity Edition Other platforms

Community Edition · Free and open source · Apple Silicon · macOS 12 or later

Hi, I’m Sparky.RigSpark’s hardware-aware sidekick.

Reads your rig.RAM, VRAM, unified memory, bandwidth and free disk. Locally.

Ranks all 254 models.A verdict and a binding reason, before a single byte downloads.

Sparky, a graphite robot with an amber visor

Three moods. One honest answer.

Yes. Slow. No.

Every model in the catalog gets a verdict sized to your machine, with the reason it was given.

Three Sparky robots: a smiling amber visor, a sleepy yellow visor and a red visor with crossed-out eyes

Try it here

Ask Sparky.

Pick your graphics, memory, free disk and a model. This preview uses RigSpark’s catalog sizes, fit rule and sourced speed data. The real CLI measures your hardware.

yes

Graphics.

Memory.

Unified memory on Apple Silicon.

Free disk.

The download has to fit before anything runs.

Model.

At the quantization rigspark up installs, default context.

Runs well.

Fits in memory with headroom.

needs 9.9 GiB18.7 GiB budget
download 8.6 GiB100 GB free

About 7–13 tok/s on Apple silicon Pro (estimate for Ollama)

$ rigspark can-run qwen3:14b → yes, exit 0

Preview only. Fit uses RigSpark’s rule: usable memory (unified or system RAM minus a 2 GiB OS reserve, or the GPU’s VRAM) with 15% headroom, and the download must fit on disk. Speed comes from RigSpark’s sourced bandwidth classes; hardware without one shows unknown rather than a guess. rigspark can-run measures your real machine.

How it works

One command ranks
every model.

rigspark recommend sizes each model’s weights and KV cache against your machine and ranks the offline catalog. Models that can’t fit are listed with the reason.

$ rigspark recommend
Ranked local LLMs for arm64/darwin (34.0 GiB ram usable):

Rank  Model              Params  Est. Mem  Verdict   Est. tok/s
   1  qwen3:30b-a3b         30B  19.9 GiB  ✓ yes    55.6–103.3
   2  kimi-vl-a3b           16B  10.7 GiB  ✓ yes    55.6–103.3
   3  qwen3:32b             32B  21.6 GiB  ⚠ slow      5.2–9.7
   4  deepseek-r1:32b       32B  21.3 GiB  ⚠ slow      5.2–9.7
   5  gemma3:27b            27B  18.6 GiB  ⚠ slow     6.2–11.5
   ⋮

Run the top pick:  rigspark up qwen3:30b-a3b

Won't fit (14):
  ❌ deepseek-r1:671b  (ram-bound)
  ❌ llama3.3:70b      (ram-bound)
Real output on an arm64 Mac with 34.0 GiB of usable RAM (quant, backend and license columns trimmed).
  1. Score. Hardware is detected locally. Unknown figures stay unknown; nothing is invented.
  2. Check. rigspark can-run qwen3:8b gives one verdict and exits non-zero only on no, so it works as a CI gate.
  3. Run. rigspark up pulls, verifies the weights and serves on 127.0.0.1. Then rigspark chat.
Real rigspark recommend output and the browser workspace. Unmute for sound.

Small robot. Big numbers.

254
catalog models
5
native runtimes
0
network calls for advice
MIT
native Rust, free forever
New in 3.0.5

A real Mac app

Install, click once, chat. llama.cpp built in.

Download for Mac
Sparky waving hello
Next release

Terminal UI

Every model, ranked for your machine.

Follow releases
RigSpark terminal UI ranking models in a sortable table with a memory fit gauge
New in 3.0.3

Ming Design UI/UX mockups

UI/UX mockups with Ming Design INT8, plus FLUX.1 schnell and Wan 2.1 through your own ComfyUI. Ming tested on a 36 GiB Apple Silicon Mac; CUDA unverified. Generated text may be garbled.

Read the changelog
Real Ming Design generation in RigSpark chat: a plant-care mobile app mockup rendered inline through local ComfyUI
New in 3.0

Workspace

Redesigned and quieter: one model chip, an inspector on demand, light and dark. Still loopback only.

What’s new
The redesigned RigSpark workspace: Sparky's welcome, a single composer and an inspector with live memory, CPU and disk metrics

AI Hardware Score

0–100. rigspark doctor names your bottleneck.

Ask Sparky
Sparky holding a processor chip

Privacy

Your hardware.
Your data. Your call.

Most “can I run it” tools guess. Sparky refuses to. Every figure is sourced, or it says unknown.

Sparky standing in a studio spotlight

  • Offline advice.

    recommend, can-run and doctor make zero network calls. Same catalog, same answer.

  • Unknown stays unknown.

    Unsourced throughput is printed as unknown. Models are still ranked by weights, never dropped.

  • Fail-closed integrity.

    up and switch check pulled weights against the catalog digest and refuse anything that doesn’t match.

  • Loopback only.

    Model servers and the workspace bind to 127.0.0.1 with host-header checks.

Which is right for you?

Ollama is an excellent runtime. RigSpark sits on top of it and three others, and tells you what to run before you run it.

A runtime alone

Ollama, llama.cpp, MLX or LM Studio

  • Run inference
  • Not included: Hardware-aware recommendations
  • Not included: yes / slow / no verdicts + est. tok/s
  • Not included: AI Hardware Score (0–100)
  • Not included: KV-cache-aware context sizing
  • Not included: Loopback-only browser workspace
  • Not included: Agents, skills & MCP tools
  • Not included: Multi-backend (4 runtimes)
  • Not included: SHA-256 integrity verification
  • Not included: Portable memory migration
  • Not included: Offline, deterministic advice

Runtime + RigSpark

Sparky decides, your runtime runs

  • Run inference
  • Hardware-aware recommendations
  • yes / slow / no verdicts + est. tok/s
  • AI Hardware Score (0–100)
  • KV-cache-aware context sizing
  • Loopback-only browser workspace
  • Agents, skills & MCP tools
  • Multi-backend (4 runtimes)
  • SHA-256 integrity verification
  • Portable memory migration
  • Offline, deterministic advice
A Sparky toy on a desk next to a laptop showing a ranked model table

Made for machines
you already own.

One command ranks every model for your rig. One more serves it on 127.0.0.1.

Community Edition · Available now

Get RigSpark.

RigSpark Community Edition is free, open source and available now for macOS and Windows. The desktop app bundles llama.cpp, so the first model runs right after install. No Node.js, Python or compiler needed.

The Mac app is signed with a Developer ID and notarized by Apple, so it opens like any other app.

If Windows SmartScreen warns about the installer: More info → Run anyway.

More downloadsCommand line for macOS, Linux and Windows, Homebrew, cargo and Docker

Command line

Homebrewbrew install shashankswe2020-ux/tap/rigspark
cargo-binstallcargo binstall rigspark-cli rigspark-gui
From sourcecargo install rigspark-cli --locked --bin llmup --bin rigspark && cargo install rigspark-gui --locked
Containerdocker pull ghcr.io/shashankswe2020-ux/rigspark:latest

Archives are unsigned. Verify them against SHA256SUMS. On macOS, clear quarantine with xattr -dr com.apple.quarantine rigspark-*. The container is an older CLI image that sees only the container’s resources.

Questions? Sparky has answers.

Figures are RigSpark’s offline estimates, not benchmarks.

How do I find out which LLMs my computer can run?

Install RigSpark and run rigspark recommend. It detects your GPU, VRAM, RAM and free disk, ranks every catalog model with a yes, slow or no verdict and estimated tokens per second, and lists the models that won’t fit with the reason. rigspark can-run <model> checks a single model.

How much VRAM do I need to run an 8B, 14B, 32B or 70B model?

At Q4_K_M and default context, RigSpark estimates about 5.3 GiB for Llama 3.1 8B, 9.9 GiB for Qwen3 14B, 21.6 GiB for Qwen3 32B and 45.5 GiB for Llama 3.3 70B. Longer contexts add KV cache on top.

Can I run local LLMs on a Mac with Apple Silicon?

Yes. Apple Silicon GPUs share unified memory with the CPU, so RigSpark sizes models against system memory and auto-selects MLX for models with an MLX build, then Ollama, then llama.cpp.

Can RigSpark generate images and videos locally?

Yes. With your own ComfyUI running on 127.0.0.1, rigspark generate image or rigspark generate video installs pinned, SHA-256 verified FLUX.1 schnell or Wan 2.1 weights and runs a built-in workflow. You can also switch the chat composer to Image or Video in rigspark gui. Paid ComfyUI Partner nodes are refused and nothing is sent to the cloud.

Does RigSpark send my data anywhere?

Recommendations, can-run and doctor make no network calls and use a verified local catalog. The explicit catalog --update action downloads a signed snapshot; normal startup does not. Model servers and the browser workspace bind to 127.0.0.1 only.

Is RigSpark the same project as llmup?

Yes. RigSpark 2.0 renamed llmup. The llmup command remains as a compatibility alias, and the crates moved from llmup-* to rigspark-*. See the changelog for moving your data directory.

RigSpark

Let Sparky check your rig.

Free and open source. MIT licensed. Native Rust.

DownloadStar on GitHubBuy me a coffee
Sparky with arms out, ta-da