mere-run

mere.run is a local-first multimodal inference runtime built around one installable CLI for Apple Silicon and headless Linux. The project covers text, code, agents, image generation, vision, speech, music, video, 3D, local API serving, and workflow graphs from the same runtime. Its core promise is practical: keep generation on machines you own when that is possible, avoid sending private creative or business data to a cloud model by default, and still expose enough commands for real production work. The README positions the CLI as the public surface and the optional macOS Studio as a companion that shares the same model store and run history. A first-run flow starts with setup, model capability inspection, and guided examples before pulling large models. The documented command groups include image generation and LoRA training, local chat and embeddings, OCR and segmentation, speech synthesis and transcription, audio restoration, music generation, video generation, OpenAI-compatible serving, and typed workflow graphs. That breadth makes mere.run different from single-purpose local model wrappers. It is closer to a local creative operating layer: one place to pull models, check memory fit, run preflights, produce outputs, serve APIs, replay plans, and move jobs across approved Macs or Linux nodes through its relay system. Builders can keep sensitive assets local while still using structured commands, plugins, progress JSON, and durable run directories that fit automation pipelines. The best users are developers, artists, VFX teams, privacy-sensitive startups, and local-AI hobbyists who already have capable hardware and want more than chat. The product is not a hosted SaaS that hides infrastructure. Model availability, memory fit, download size, and licensing vary by command, so the docs repeatedly point users to capability checks and model-source notes before large downloads or redistribution. Pricing is not presented as a conventional SaaS plan on the GitHub source. The repository and docs emphasize local execution and signed releases, so OpenTools lists the core project as free/open-source pending further commercial packaging details. The main operational cost is user-owned hardware, electricity, storage, and any optional relay or plugin workflow the user chooses to run. The tradeoff is that local-first tools ask users to care about hardware. mere.run helps by exposing capability checks, preflight commands, model guides, checksums, progress JSON, and run directories, but builders still need to understand which models fit on a given machine. That is a feature for the right audience. A creative technologist can keep client assets local, automate repeatable generation jobs, and serve local APIs to adjacent tools. A privacy-sensitive startup can prototype multimodal workflows without sending every input to a hosted provider. A hobbyist can learn how image, speech, video, and agent workflows differ while keeping them under one command surface. If you want a managed cloud product, this is likely too hands-on. If you want control, it is compelling.