v2.2.0 Signed catalog updates and richer chat →

Hardware-aware local inference

Run the right local model. Before you download.

RigSpark scores your hardware and gives every model a yes, slow or no verdict with estimated tok/s before you download. Then it verifies, serves and chats with your pick.

cargo binstall rigspark-cli rigspark-gui

Download for your OS Watch demo

69: curated models
4: native runtimes
0: network calls for advice

Real `rigspark recommend` output and the browser workspace. Unmute for sound.

yes

Fits in memory and meets bandwidth. Shows an estimated tok/s range.

slow

Fits, but memory bandwidth limits speed. Still shows the estimate.

Won’t fit. You see the binding reason, so you can skip the download.

One engine, three views

See the decision, not just the dashboard.

Real RigSpark screens, from hardware fit to verified local chat.

RigSpark terminal UI ranking local models in a sortable table with memory need, verdict, tok/s and score beside a memory fit gauge

RigSpark model details with recommendation score, hardware fit, throughput evidence and memory sizing

RigSpark local browser workspace ready to chat with a selected model

New in 2.2.0

Update the catalog without replacing RigSpark

Catalog updates are explicit, signed and activated atomically. Normal startup and recommendations stay offline, while failed or incompatible updates leave your current verified snapshot untouched.

Visible provenance. Inspect the active source, revision, digest and model count in the CLI or Models view.
New model families. Bonsai and Qwen3.5 join the curated catalog; unsourced throughput remains unknown.
Richer local chat. Pinned KaTeX renders TeX offline behind the existing sanitized Markdown boundary.

Install 2.2.0 Read the changelog

Update from the Models view, or run rigspark catalog --update.

How it works

One command ranks every model for your machine

rigspark recommend detects your RAM, GPU and free disk, sizes each model’s weights and KV cache, and ranks the offline catalog. Models that can’t fit are listed separately with the reason, and the top pick is one command away.

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.

rigspark — zsh

$ 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).

Terminal UI Next release

A keyboard-first terminal UI, not just pretty text

Run rigspark with no arguments and a native terminal UI opens when your terminal is at least 60×16. Sort and filter a ranked table, read memory fit at a glance, and compare models side by side, by keyboard or mouse. Accessible mode (--accessible) serves screen readers, and CI or piped output falls back to plain text automatically.

rigspark terminal UI filtering the model table to 18 qwen models with matched characters highlighted — **Fuzzy search with highlighted matches.** Type to narrow the table; matched characters are underlined, and verdict filters combine with search. Click a row to select it, click again to open it, or click a column header to sort. `/` search `Ctrl+U` reset `S` reverse sort `Y` copy `rigspark up`

Everything in one tool

From hardware check to local chat

AI Hardware Score

A 0–100 score names your bottleneck, and every catalog model gets a yes / slow / no verdict with est. tok/s before any download.

Context-window sizing

KV-cache-aware memory math with GQA-correct attention geometry. Re-rank at --context N, find each model’s ceiling with --max-context, or size a quantized cache with --kv-cache.

Portable memory

Chats are recorded to local memory. rigspark migrate --from qwen3:8b --to qwen3:14b carries it to a bigger model; preview with --dry-run.

Agents & skills

Reusable personas and instruction blocks, stored locally as markdown. Bundle skills into an agent or toggle them per message.

MCP connectors & tools

Attach Model Context Protocol servers, local stdio or loopback HTTP/SSE. The model calls enabled tools in an agentic loop, with your approval.

Inline images & graphs

Let an agent run code in your workspace; generated images and graphs render in the chat, served from a loopback-only artifacts endpoint.

Signed catalog updates

Explicitly download and atomically activate a verified catalog without replacing the app. Revision rollback and verified fallback protect advice.

Offline TeX rendering

Pinned KaTeX assets render equations locally while malformed or unsafe rich responses stay behind the sanitized Markdown boundary.

Evidence-backed models

Bonsai, Qwen3.5 and Qwen3.6 carry pinned provenance. Unsourced throughput and attention geometry remain visibly unknown.

Keyboard-first TUI Next

Sort any column, filter by verdict, fuzzy-search names and compare up to four models side by side. Mouse, light and high-contrast themes (RIGSPARK_THEME) included.

Streaming terminal chat Next

Replies stream as they arrive with headings, lists and code rendered. Scroll back with PgUp, copy the last reply with Ctrl+Y, and see first-output and total time.

Live lifecycle checklist Next

up, switch and down show each observed stage as done, failed or unconfirmed, with elapsed time. Stage completion never claims overall success.

Browser workspace

Pick a model that fits, start it, chat with it

rigspark gui opens a local-first workspace on 127.0.0.1 that uses the same engine as the CLI: the same verdicts, the same verified up lifecycle. Local chat stays on your machine; cloud harnesses and external tools send data only to their own providers.

Models view ranking recommended models with Runs well and Runs slowly verdicts — Recommended models ranked for your hardware, with the same verdicts and est. tok/s as the CLI. **Start** brings one online through the verified lifecycle.

Approving a read-only MCP tool call and rendering the result in chat — Two read-only WHOOP tools are proposed, explicitly approved, and used to build a live health briefing. Connectors are local `stdio` commands or loopback HTTP/SSE; enable, disable and inspect each one’s tools.

loopback-only
host-header checked
same engine as the CLI
pluggable harnesses

Built to be trusted

Honest numbers, local by default

Offline advice

recommend, can-run and doctor make no network calls. Advice comes from a curated, cited catalog, so it is reproducible.

Unknown stays unknown

When a figure can’t be sourced, RigSpark prints unknown. Models are still ranked by weights, never silently dropped. Estimates are not benchmarks.

Fail-closed integrity

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

Loopback only

Model servers and the browser workspace bind to 127.0.0.1. Nothing is exposed to your network.

Curated catalog

69 models. Evidence attached.

The catalog is an offline, reviewable dataset, not a scrape. Each entry ties a model to its architecture, context, quantizations, artifact size, integrity digest and official sources. Activate independently published snapshots with catalog --update, or use --kv-cache q8_0 or q4_0 to re-size advice for quantized KV caches.

69: curated models
4: runtimes
SHA-256: artifact integrity
Weekly: coverage audit

Curate Architecture, license, capabilities, context, quantization and official sources are reviewed and pinned.
Bootstrap The snapshot deterministically generates the same offline advice dataset on every machine.
Enrich Known Ollama artifacts gain exact model-layer bytes and content digests without rewriting curated facts.
Audit Freshness and upstream coverage produce review candidates, never automatic admissions.

Honesty gate: missing attention geometry stays unknown. Coverage checks can’t enumerate missing tags inside a model already in the catalog.

Backends

Four runtimes, one workflow

OllamamacOS, Linux, Windows · recommended defaultFull lifecycle
llama.cppmacOS, Linux, Windows · GGUF from Hugging FaceFull lifecycle
MLXApple Silicon · preferred when installedFull lifecycle
LM StudiomacOS, Linux, Windows · you run the serverAttach-only

Auto-selected for your platform and installed runtimes; override with --backend. All bind 127.0.0.1 only.

Chat with any runtime from the workspace

A backend picker reaches every runtime directly: the same model chatting locally through each one.

Chatting with a model served by Ollama — Ollama · 127.0.0.1:11434

Chatting with a model served by llama.cpp — llama.cpp · owned server

Chatting with a model served by MLX on Apple Silicon — MLX · Apple Silicon

Chatting with a model served by LM Studio — LM Studio · attach-only

Comparison

RigSpark vs. Ollama

Ollama is an excellent runtime. RigSpark adds a hardware-aware layer on top of it, and of three other runtimes.

Feature	Ollama	RigSpark
Run inference	✓	✓
Hardware-aware model recommendations	–	✓
yes / slow / no verdicts + est. tok/s	–	✓
AI Hardware Score (0–100)	–	✓
Context-window sizing (KV-cache aware)	–	✓
Interactive terminal UI	–	✓
Browser workspace (loopback-only)	–	✓
Agents & skills library	–	✓
MCP connectors & tools	–	✓
Inline images & graphs in chat	–	✓
Accessible mode (screen readers)	–	✓
Multi-backend (4 runtimes)	–	✓
SHA-256 integrity verification	–	✓
Portable memory migration	–	✓
Offline deterministic advice	–	✓

Install

One archive, three native binaries

Each release ships rigspark, the llmup alias and the rigspark-gui workspace. No Node.js, Python or compiler needed: unpack, add the folder to your PATH, run.

macOSApple Silicon macOSIntel Linuxx64 LinuxARM64 Windowsx64

Archives are unsigned. Verify them against SHA256SUMS. On macOS, clear quarantine after unpacking with xattr -dr com.apple.quarantine rigspark-*.

Homebrew, on macOS or Linux

brew install shashankswe2020-ux/tap/rigspark

Prebuilt, with cargo-binstall

cargo binstall rigspark-cli rigspark-gui

From source, with Cargo

cargo install rigspark-cli --locked --bin llmup --bin rigspark && cargo install rigspark-gui --locked

Container: an older CLI image that sees only the container’s resources

docker pull ghcr.io/shashankswe2020-ux/rigspark:latest

Commands

The full local-LLM lifecycle

Every command supports --json and --help. Use --accessible for screen readers.

recommend: Default. Ranks the catalog for this machine; filter with --task, size context with --context N.
can-run: One verdict with the binding reason and est. tok/s. Exits non-zero only on no.
doctor: AI Hardware Score, bottleneck diagnosis, backend status and environment checks.
catalog: Browse models, inspect provenance with --status, or explicitly activate a signed snapshot with --update.
up: Pull with SHA-256 verification, then serve on 127.0.0.1.
chat: Terminal chat with streaming, multi-line input (Ctrl+J) and session memory.
gui: Open the loopback-only browser workspace.
ls: Show the active model and server endpoint.
switch: Change the active model without re-pulling, with a keyboard picker in the TUI.
migrate: Move conversation memory to another model; preview with --dry-run.
down: Stop the local server cleanly.

FAQ

Questions before you run models locally

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.

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.

Know before you download.

Free and open source under the MIT license.

Install RigSpark Star on GitHub