// AI TOOLS · MERE.RUN

Mere.run

An open-source, local-first creative AI runtime with a single CLI that generates text, images, video, music, sound, speech, and 3D on your own Apple Silicon Mac or Linux box — models pulled into a local store and run fully offline, no Python to manage.

Last verified · 2026-07-31 · by Moe Ameen

What Mere.run is

Mere.run is an open-source, MIT-licensed "creative AI runtime" that runs on your own hardware. The idea is a single command-line tool — mere.run — that generates across every media type from one interface: text and code, images, short video, music and sound effects, speech, and 3D geometry, all executing locally with nothing sent to a cloud service. It surfaced publicly as a Show HN with the pitch "local text, image, video, music and 3D from one CLI, no Python." The "no Python" part is the design statement: you install a signed binary or a small package (a macOS .dmg, a one-line installer, an npm package, or a Linux build), not a Python environment full of dependencies to break.

Under the hood it is built in Swift and leans on Apple's MLX framework for on-device inference, plus a vendored llama.cpp for text. On macOS it ships both the CLI and an optional SwiftUI desktop app (MereRun.app) with a studio interface; on Linux it's a headless CLI. It's developed and validated on Apple Silicon macOS 15 or newer and on headless Linux (x86_64, with arm64 builds aimed at high-memory NVIDIA systems). You pull models into a local model store — e.g. `mere.run model pull …` — and then generate offline; there's no account, API key, or per-token bill.

The model catalog rotates and spans families rather than a single fixed lineup: text models like Gemma and Qwen, image models like Krea 2, HiDream, and Z-Image, video models like LTX and Wan, music models like ACE-Step and Magenta RealTime, speech via on-device ASR/TTS, and vision models like SAM for masking and detection. Beyond one-shot generation it also exposes an OpenAI-compatible local API server (`mere.run api serve`) for chat, embeddings, images, TTS, and STT, so it can act as a private, drop-in backend for your own apps.

Two honest points. First, this is a developer-facing, hardware-bound tool, not a consumer app: which models you can run and how fast depends on your machine, with heavier media models wanting large amounts of unified memory (the docs describe memory bands running from roughly 16 GB up past 96 GB). Free software still assumes you own capable hardware. Second, like every generation tool, it stops at producing a file — it renders a clip, an image, a track, a mesh. It has no concept of a hook frame, feed-ready captions, a vertical crop, a brand template, a schedule, or a publish target. That gap is exactly where a distribution layer takes over.

What you can make with it

  • Text, chat, and code generation — plus embeddings — from local language models
  • Images from prompts, with editing, LoRA training, and 2D-to-3D reconstruction
  • Short AI video clips and animation, generated on-device
  • Music tracks and sound effects from local audio models
  • Speech: text-to-speech, transcription, and speaker diarization from voice models
  • 3D and geometry outputs — depth maps, point clouds, and meshes
  • A private, OpenAI-compatible local API endpoint to wire into your own apps

How Kompozy turns Mere.run output into content

Mere.run's whole point is that everything generates locally and privately — which also means everything lands as a bare file in a folder on your disk. A rendered clip has no captions for muted autoplay, no vertical crop for a feed, no branded frame, and no way to reach an audience; a generated image is one PNG, not a Carousel; a music track is a WAV with nowhere to go. Kompozy is the cloud finishing-and-distribution layer that turns those local outputs into on-brand, scheduled posts across every platform, and adds the net-new formats a raw runtime can't produce.

The concrete workflow: generate your assets in mere.run — a video from LTX or Wan, an image from Krea 2, a music bed from ACE-Step, a transcript from the ASR model — then drop them into Kompozy as sources. Clipped Shorts cuts a longer video into vertical 9:16 segments with word-synced captions burned in for TikTok, Reels, and Shorts. A locally generated image becomes a brand-exact Carousel or Photo Post via HyperFrames; a track can score a Listicle Video or Marketing Short. And because Kompozy generates as well as finishes, it can wrap the same idea in formats mere.run has no model for at all — Persona Shorts and HeyGen avatar video that put a consistent face on camera, Quote Graphics, a Blog Article, and an Email Newsletter — all governed by your Persona Brief so a batch made from local files still sounds like one voice. Autopilot and a per-post review pipeline then schedule and publish the set across eight social platforms plus blog and email. Mere.run keeps generation private and offline; Kompozy makes what you generated finished, branded, and everywhere.

  1. Pull the models you need into mere.run's local store and generate offline — video, image, music, or a transcript.
  2. Bring the exported files into Kompozy as sources for a format — Clipped Shorts for vertical cuts, Carousel or Photo Post for images.
  3. Let Kompozy apply your Persona Brief and HyperFrames brand styling, and add formats mere.run can't — Persona/HeyGen avatar video, Quote Graphics, a blog, a newsletter.
  4. Fan the same idea into multiple formats at once instead of hand-editing each platform.
  5. Schedule and publish across the eight social platforms plus blog and email using the review pipeline or autopilot.

Frequently asked questions

What is Mere.run?

Mere.run is an open-source, MIT-licensed local-first "creative AI runtime" — a single command-line tool that generates text, images, video, music, sound, speech, and 3D on your own hardware, fully offline. It's built in Swift on Apple's MLX framework (plus vendored llama.cpp), runs on Apple Silicon macOS and headless Linux, and ships an optional macOS studio app.

Is Mere.run really local and offline?

Yes. You pull models into a local model store and inference runs on your own machine — no account, API key, or cloud call, so nothing you generate leaves your hardware. That privacy is its main draw over cloud generators. The trade-off is that you supply capable hardware; heavier media models want large amounts of unified memory.

What does "no Python" mean for Mere.run?

It means you don't have to install and maintain a Python environment to run it. Mere.run installs as a Swift binary or package (with a macOS .dmg and Linux CLI builds), so there's no dependency stack to break — a contrast to the Python-heavy setup most local AI tools require.

Can Mere.run publish content to social media?

No. Mere.run generates media files locally; it has no concept of captions for a feed, brand templates, scheduling, or publishing. To turn its outputs into finished, on-brand posts and distribute them, pair it with a content engine like Kompozy, which captions, reframes, brands, and publishes across nine destinations — the eight social platforms plus blog and email.

How is Mere.run different from a cloud tool like Pollo AI or Venice AI?

Pollo AI and Venice AI route many models through a hosted cloud interface; Mere.run runs the models on your own machine with no server in the loop. You trade cloud convenience and instant scale for privacy, no per-token cost, and full offline control — at the price of owning the hardware to run it.

Related tools

  • NativA private, on-device AI app for Apple devices that runs open-weight language models and image models locally — no accounts, no subscription, and no internet required once a model is downloaded.
  • LM Studio BionicLM Studio's AI agent built for open models — a local-first assistant that inspects and edits code, works over your documents, and runs on models you download, connect, or call in the cloud.
  • Gemma 4 26B Local EngineRunning Google's Gemma 4 26B model on your own machine — via Ollama, llama.cpp, MLX, or vLLM — for free, private, offline text drafting on consumer hardware.
  • TurboFieldfareAn open-source Swift + Metal engine that runs Google's Gemma 4 26B model on any Apple Silicon Mac in about 2 GB of RAM — by streaming only the experts each token needs from the SSD instead of loading the whole model.
  • Venice AIPrivacy-first AI platform that routes 200+ models — text, image, audio, and video — without storing your data.
  • Pollo AIAn all-in-one AI creative suite — 100+ video and image models plus purpose-built apps for avatars, UGC ads, and product video, in one workspace.

← All AI tools · Get started →