# Muse Spark Open: Meta releases an open-weight model, if you have a terabyte of memory

URL: https://ai-pulse.aglaganova.com/blog/muse-spark-open-meta-releases-an-open-weight-model-if-you-have-a-terabyte-of-memory
Published: 2026-10-21
Author: Roman Aglaganov, Full-stack developer
Category: Models
Region: Global
Tags: open-weights, coding, agents

> Meta’s new open-weight model has 816 billion parameters and a million-token context. It runs on Macs, but only on two top-end Mac Studios linked together.

![Illustration: a dense oval cloud of grey and black dots above two linked boxes joined by a red dot](https://ai-pulse.aglaganova.com/api/media/file/muse-open-1.png)

Meta released Muse Spark Open on 15 October, a model it calls “the most capable US open-weight model to date”. It follows Muse Wonder, Meta’s closed flagship, and answers a question many developers had asked since: whether Meta was still releasing models anyone can download.

## What it is

**Size:** 816 billion parameters, of which 42 billion are active for each token (a mixture-of-experts design), with a context window of one million tokens.

**Purpose:** it is natively multimodal and built for coding and agent work: long tasks, software engineering and multi-step reasoning.

**Where to get it:** the weights are on Hugging Face and the big clouds (Amazon SageMaker, Google Cloud, Microsoft Foundry), and it is served by API providers such as Together AI, Fireworks, Groq and OpenRouter.

**The licence:** it uses Meta’s own Muse Model License, which Meta says has “limited commercial terms” that let many developers use it for free. It is not an open-source licence like Apache 2.0.

## Running it on a Mac

Meta showed the model running locally on two Mac Studios with Apple’s M5 Ultra chip, linked so that they “behave like one supercomputer with a terabyte of unified memory”. Engadget noted that means at least 1 TB of RAM. In demos, Meta’s Muse Code terminal used it to make changes to Ice Cubes, an open-source Mastodon client, and to find and crop footage in DaVinci Resolve.

Meta says the model beats other popular open-weight models, but those are its own evaluations. It also says the model received extra safety training because open weights cannot be recalled, and that it passed the release thresholds of Meta’s Superintelligence Scaling Framework; a full evaluation report is promised.

## Why it matters

Meta was once the biggest name in open models, and keeping Muse Wonder closed made people wonder if that era was over. Muse Spark Open shows it is not, though Meta’s best model stays closed. At this size, “runs locally” means a workstation costing as much as a car, not a laptop. For most developers, the practical win is a strong model they can host where they choose, fine-tune and inspect.

## Sources

- [Meta AI: Muse Spark Open](https://research.meta.ai/blog/muse-spark-open)
- [TechCrunch: Meta launches an open-weight AI model that runs on Macs](https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/)
- [Engadget: Meta built a huge open-weight AI model for Macs, but you’ll need 1TB of RAM to run it](https://www.engadget.com/2287833/meta-built-a-huge-open-weight-ai-model-for-macs-but-youll-need-1tb-of-ram-to-run-it/)