RigSpark

Product Introduction

1. Definition: RigSpark is a free, open-source (MIT licensed) local AI orchestration desktop application and command-line tool written natively in Rust for macOS, Windows, and Linux. Categorized under local LLM managers and AI developer tooling, it interfaces directly with five distinct runtimes—Ollama, llama.cpp, MLX, LM Studio, and BitNet—while integrating with local ComfyUI instances to serve text generation, image creation, short video synthesis, and voice dictation entirely offline.

2. Core Value Proposition: RigSpark exists to eliminate the guesswork of running open-source AI models on consumer hardware by automatically auditing system RAM, GPU specifications, and available disk space. Using its proprietary hardware-aware evaluation engine, the tool categorizes models with clear "yes," "slow," or "no" verdicts accompanied by estimated token-per-second (tok/s) throughput before initiating any downloads. By prioritizing local execution over cloud dependencies, RigSpark ensures absolute data privacy with zero accounts, zero meters, and zero telemetry.

Main Features

1. Hardware-Aware Recommendation Engine (rigspark recommend & can-run): RigSpark scans local system resources—detecting CPU architecture, available VRAM, system RAM, and storage limits—to size model weights and KV caches with GQA-correct geometry. Every model in its 254+ item catalog (plus user-supplied GGUF files from Hugging Face) receives a deterministic verdict ("yes" for optimal performance, "slow" for memory bandwidth bottlenecks, or "no" for hardware incompatibilities) along with estimated execution speeds. The can-run command can be used in Continuous Integration (CI) pipelines to exit non-zero on unsupported hardware configurations.

2. Multi-Runtime and Multi-Modal Orchestration: RigSpark unifies five distinct inference runtimes (Ollama, llama.cpp, MLX for Apple Silicon, LM Studio, and BitNet for CPU execution) under a single command-line and browser workspace interface. Beyond large language models (LLMs), the platform orchestrates multimodal workflows including text chat and code generation, FLUX.1 schnell and Ming Design mockups for image generation via ComfyUI, Wan 2.1 and Wan 2.2 for text-to-video clips, and speech-to-text/text-to-speech pipelines using Whisper, Kokoro, and Chatterbox.

3. Fail-Closed Integrity and Privacy-First Loopback Architecture: Security and data integrity are enforced through strict local constraints. Model weights pulled via rigspark up or rigspark switch undergo mandatory SHA-256 digest verification against the catalog snapshot, refusing any corrupted or unverified files. All local model servers, model catalogs, and browser workspaces bind strictly to 127.0.0.1 with mandatory host-header verification. Commands like recommend, can-run, and doctor make zero network calls, operating entirely offline on signed catalog updates.

Problems Solved

1. Pain Point: Users frequently encounter trial-and-error bottlenecks when attempting to run local AI models, downloading multi-gigabyte weight files only to discover their GPU VRAM or system RAM is insufficient, resulting in out-of-memory (OOM) crashes or unacceptably slow inference speeds. Additionally, standard inference runtimes lack hardware diagnostic scoring, context-window memory math, and cross-backend management.

2. Target Audience: Local AI enthusiasts, machine learning engineers, privacy-conscious developers, software engineers, and power users seeking to run open-source models (such as Llama, Qwen, DeepSeek, and Gemma) locally without third-party API exposure or cloud subscription meters.

3. Use Cases: Running offline LLMs for confidential data processing and secure coding assistants; generating design mockups, UI layouts, and multimedia content locally via ComfyUI integrations; deploying local speech transcription and text-to-speech agents; and automating hardware compatibility checks within local development pipelines.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Strengths: Native Rust performance with zero runtime bloat (no Python or Node.js required for desktop installers); deterministic, offline hardware compatibility audits; support for 5 distinct backends; SHA-256 integrity verification; Model Context Protocol (MCP) tool integration; and portable chat memory migration (rigspark migrate).
   • Limitations: Advanced multimodal features like short video generation (Wan 2.1/2.2) and local image pipelines require an external, properly configured local instance of ComfyUI; Windows code signing relies on community-tier infrastructure (SignPath.io); and raw token estimates are heuristics rather than guaranteed real-time benchmarks.

2. Key Alternatives & Differentiation:

   • Ollama / LM Studio: While excellent standalone runtimes, they primarily focus on model execution rather than cross-runtime management, hardware scoring, and granular pre-download verification. RigSpark acts as an orchestration layer on top of Ollama, llama.cpp, and others, providing algorithmic hardware matching (yes/slow/no verdicts) that standalone runtimes lack.
   • Custom Python Scripts: Manual orchestration scripts lack unified desktop workspace interfaces, signed catalog integrity checks, loopback security hardening, and native multi-backend lifecycle management (up, switch, down).

Frequently Asked Questions (FAQ)

1. How does RigSpark determine if my computer can run a specific AI model? RigSpark executes local hardware profiling to read available system RAM, GPU VRAM, and free disk space. It then calculates the model's weight memory footprint and KV cache requirements using GQA-correct geometry, comparing them against your machine specs to output a verified "yes," "slow," or "no" verdict with estimated tokens per second.

2. Does RigSpark require an internet connection or send my data to the cloud? No. Diagnostic commands such as rigspark recommend, rigspark can-run, and rigspark doctor operate entirely offline with zero network requests. Model inference servers and the browser workspace bind exclusively to 127.0.0.1 (loopback), ensuring your chat data, prompts, and personal files never leave your machine.

3. Can I use GGUF files downloaded directly from Hugging Face? Yes. Starting with version 3.0.4, RigSpark allows you to paste any Hugging Face GGUF repository link into the tool, letting you run custom or newly released models directly through your preferred supported backend.

4. How do local image and video generation features work in RigSpark? Image and video generation capabilities run through your own local instance of ComfyUI hosted on 127.0.0.1. When you execute commands like rigspark generate image or switch the GUI composer to image/video mode, RigSpark installs and verifies pinned, SHA-256 checked weights (such as FLUX.1 schnell or Wan 2.1) and executes them locally without paid partner nodes or cloud egress.