Show HN: Local text, image, video, music and 3D from one CLI, no Python

A new command-line interface named mere.run has been developed to facilitate local execution of various AI inference tasks including text, image, video, music, and 3D processing. Designed primarily for Apple Silicon devices and headless Linux environments, mere.run aims to maximize the use of local computing resources rather than relying on cloud-based services.

This innovative runtime consolidates multiple AI modalities into a single CLI tool, enabling users to run complex workflows involving chat interactions, multimedia generation, and API serving without internet dependency. The solution supports diverse data formats such as speech, music, sound effects, video, and 3D models, thereby addressing a broad spectrum of creative and analytical applications.

On macOS, mere.run offers an optional SwiftUI-based studio application that integrates seamlessly with the CLI. This studio environment preserves command histories, model management, and runtime logs while providing native system features like Quick Look previews for generated artifacts. The macOS app is distributed as a signed DMG containing the CLI, runtime assets, and optional skill modules, streamlining installation and usage for creative professionals and developers alike.

Linux users can access mere.run as a headless package optimized for ARM64 and x86_64 architectures with CUDA acceleration. However, the Linux offerings focus solely on the CLI and do not include the graphical studio or macOS-specific packaging. Installation on Linux platforms requires specific dependencies such as CUDA Toolkit headers and libraries to enable GPU-accelerated computations.

Mere.run leverages the Hugging Face model repository for managing AI models locally. It intelligently caches and links model weights to minimize redundant downloads and storage, allowing users to share cached content across different projects. The runtime enforces a local-first security posture by default, requiring users to opt into more permissive modes that could compromise this safety.

The project draws inspiration from Python-based local inference communities, translating their approaches into a Swift-native implementation that runs efficiently on Apple Silicon hardware. It emphasizes a modular architecture with relay and node components designed to distribute workloads across multiple machines if needed, though it does not currently offer a hosted backend service.

Developers interested in contributing or customizing mere.run can access its open-source repository on GitHub, which includes documentation, build scripts, and plugin frameworks. The tooling supports validation workflows using standard utilities such as SwiftLint, ripgrep, Node.js, and pnpm, ensuring code quality and maintainability.

While mere.run currently supports a wide array of AI tasks and media types, details about specific model capabilities, licensing, and memory requirements vary by command. Users are advised to consult the official documentation and model source files before deploying large models or redistributing artifacts.

In summary, mere.run represents a significant step forward in enabling offline, local-first AI inference for multimedia workflows on modern hardware. By consolidating numerous modalities into a single CLI and complementing it with an optional macOS studio, the project offers a robust platform for artists, musicians, and developers seeking privacy, performance, and control over their AI-powered creative processes.