Overview
RigSpark is a native Rust tool for choosing and running local AI models on existing hardware. Its Sparky assistant reads RAM, VRAM or unified memory, memory bandwidth, and available disk locally, then ranks a curated catalog of 69 models before model weights are downloaded.
Each model receives a yes, slow, or no verdict. Models that fit also receive an estimated tokens-per-second range, while rejected models include the limiting reason. Estimates are explicitly presented as offline estimates rather than benchmarks.
Key Features
- Sizes model weights and context-dependent KV cache against detected hardware.
- Ranks models and preserves unknown values instead of inventing throughput figures.
- Provides an AI Hardware Score and identifies hardware bottlenecks.
- Verifies downloaded weights against catalog digests and rejects mismatches.
- Supports Ollama, llama.cpp, and MLX lifecycles, plus attach-only use with LM Studio.
- Keeps model servers and its browser workspace bound to the local loopback interface.
Local Model Workflow
The recommend command ranks catalog models for the current machine and explains models that cannot fit. The can-run command evaluates one model and returns a nonzero exit status only for a no verdict, allowing it to serve as a CI gate.
The up command downloads, verifies, and serves a selected model locally, after which chat is available in the terminal or browser workspace. Lifecycle commands expose each stage as completed, failed, or unconfirmed, and portable chat memory can be previewed and migrated to a larger model.
Agents, Tools, and Media
RigSpark stores reusable agents, personas, and instruction blocks locally as Markdown. It can connect to local standard-input/output or loopback MCP servers, with tools requiring user approval, and its chat can render agent-produced charts and images.
Version 2.3 also supports local image and video generation through a user-operated ComfyUI instance on loopback. Built-in workflows use pinned, SHA-256-verified FLUX.1 schnell or Wan 2.1 weights, and paid ComfyUI partner nodes are refused.
Privacy and Availability
Hardware recommendations, single-model checks, and diagnostics run without network calls using a verified local catalog. Catalog updates are explicit, signed snapshots can be activated or rolled back, and normal startup does not fetch an update.
RigSpark is free, MIT-licensed software distributed for macOS, Linux, and Windows. It is available as native archives, through Homebrew or cargo-binstall, from source with Cargo, and as a container whose hardware view is limited to the container’s assigned resources.