Meta Launches Muse Spark Open, an 816 Billion Parameter AI Model Built to Run on Macs

TL;DR

• Meta introduced Muse Spark Open, describing it as the most capable open-weight AI model in the United States.
• The model features 816 billion parameters, a 1 million token context window and native multimodal capabilities.
• Local deployment on Macs reportedly requires at least 1 terabyte of RAM, creating a substantial hardware barrier.

Meta Expands Its Open-Weight AI Lineup

Meta has launched Muse Spark Open, a large open-weight model designed for complex coding, agent-based workflows and local deployment on Apple computers.

The company described the release as the “most capable” open-weight model in the United States. Evaluation results published by Meta on Thursday indicate stronger performance than several competing open-weight models, according to the company.

Muse Spark Open contains 816 billion parameters and supports a context window of 1 million tokens. That capacity allows it to process substantial amounts of information within a single interaction, potentially supporting work involving extensive codebases or lengthy documents.

The model is also natively multimodal, extending its capabilities beyond text-based tasks.

Running Muse Spark Open Locally Requires Substantial Memory

Compatibility with Macs is a prominent part of the launch, although the model’s hardware demands limit its accessibility.

According to Engadget’s reporting cited in the announcement coverage, running Muse Spark Open locally requires at least 1 terabyte of RAM. The model is expected to perform best on a Mac Studio equipped with Apple’s M5 Ultra chip.

Those requirements place local deployment beyond typical consumer hardware. Developers and organizations considering an on-device setup would need to weigh the model’s capabilities against the cost and availability of suitable equipment.

API access provides another route for users who want to work with the model without maintaining the hardware needed to run it themselves.

Muse Code Demonstrations Highlight Agentic Coding Tasks

Meta paired Muse Spark Open with Muse Code, its AI coding terminal, to demonstrate workflows involving multiple development steps.

In one example, the company assigned the model work on a Mastodon client, including bug fixes, new features and a complete build.

The demonstration illustrates Meta’s focus on agentic coding: systems that can carry out sequences of development tasks rather than simply suggest individual snippets.

However, the examples remain company-reported demonstrations. Their results do not establish how consistently the model will perform across different projects, software environments or levels of task complexity.

Meta Shows Video-Editing Capabilities in DaVinci Resolve

The company also demonstrated Muse Spark Open working with DaVinci Resolve, the professional video-editing application.

Meta said the model could review and crop footage, highlighting potential uses beyond software development.

The example points toward broader desktop assistance, where AI systems interact with applications to complete practical tasks. Together with the coding demonstrations, it positions Muse Spark Open as a model intended to support work across different tools.

Open Weights Give Developers More Deployment Options

By releasing the model’s weights, Meta gives developers greater flexibility over deployment and customization than services available exclusively through proprietary APIs.

Local operation can provide more control over infrastructure and data handling. Depending on the workload and hardware costs, it may also offer economic advantages, although a model of this size requires substantial resources.

Open weights do not automatically mean unrestricted use or full transparency into training data and development methods. Those details depend on the accompanying licence and documentation.

Meta said Muse Spark Open is available now through multiple API platforms, giving users a way to explore its capabilities alongside the more demanding option of running it locally.