Meta Releases Muse Spark Open, a 816B-Parameter AI for Desktop
Meta has released the weights for Muse Spark Open, a powerful multimodal AI model optimized for Apple Silicon.
Meta AI Research has released the full model weights for Muse Spark Open, positioning it as the most capable US open-weight AI model to date. This release marks a significant shift toward running powerful, agentic AI models on personal desktop hardware, specifically Apple's latest silicon, rather than solely in the cloud.
What is Muse Spark Open?
According to the official announcement, Muse Spark Open is a natively multimodal model with a total of 816 billion parameters, of which 42 billion are active per inference. It features a 1-million-token context window, making it suitable for processing extremely long documents or codebases. Meta states the model is specifically built for complex software engineering, multi-step reasoning, and long-horizon agentic tasks. The weights are being released under the Muse Model License, which allows many developers to use it for free under limited commercial terms.
Why is running it on Apple Silicon notable?
A key claim from Meta is that Muse Spark Open is the first model in its class designed to run on a user's desktop using two Apple M5 Ultra Mac Studios. The company explains that Apple's unified memory architecture is crucial, allowing the M5 Ultra's integrated GPU to access up to 512 GB of memory per Mac Studio. With two machines, this provides enough memory to hold the entire model. The chip's high memory bandwidth (1.2 TB/s) enables fast token generation. Meta demonstrated on-device workflows, including shipping real iOS code changes to an open-source Mastodon client and driving video editing in DaVinci Resolve.
Where is the model available?
The model weights are available for immediate download from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. For those not wanting to run it locally, hosted inference is available through API partners including Together AI, Fireworks AI, Groq, OpenRouter, Baseten, Modal, and Nebius. Developers can run the model using popular inference engines like vLLM, TensorRT-LLM, llama.cpp, and Apple's MLX framework, which is optimized for Apple Silicon.
What are the safety considerations for an open-weight model?
Meta acknowledges that releasing open-weight models carries distinct considerations. The company states that Muse Spark Open has undergone additional safety post-training as a result. Its release was evaluated under the Meta Superintelligence Scaling Framework, and it met the required thresholds for severe risks. Detailed safety evaluation results are provided in the model card, with a full report promised soon. The company frames open-weight releases as part of its approach to advancing safety work across the AI field.
How does this relate to work in Grafy?
For developers and creators working in Grafy, the release of powerful open-weight models like Muse Spark Open expands the frontier of what's possible with local, high-fidelity AI. While Grafy itself provides cloud-based access to hundreds of models via OpenRouter and other providers in apps like Grafy Chat, the local deployment option is significant. Developers using Grafy Code, our AI coding studio, can explore similar complex, multi-step coding agent workflows. The ability to run such a large model locally also complements research workflows in Reader, where long-context, reasoning-heavy models are invaluable for interrogating large PDFs and technical papers.
Frequently asked questions
What does "open-weight" mean?
Open-weight means the model's trained parameters (the "weights") are publicly released for download and local execution. This contrasts with closed API models where users only access outputs. It offers greater control, privacy, and customization but requires significant computational resources to run.
Can I run Muse Spark Open on a single Mac Studio?
According to Meta's specifications, running the full 816B-parameter model requires two Apple M5 Ultra Mac Studios linked together to provide the necessary unified memory (up to 1 TB combined). The model is optimized for this two-system setup using the MLX framework.
What are "agentic capabilities"?
Agentic capabilities refer to an AI model's ability to perform long sequences of actions or reasoning steps autonomously to achieve a goal. For Muse Spark Open, this includes tasks like writing code, running tests, fixing bugs, and editing video clips end-to-end without constant human intervention.
More news
Meta Launches Muse Spark Open, a Powerful AI Model for Apple Macs
Meta's Muse Spark Open, an 816B-parameter open-weight model optimized for Apple silicon.
Meta releases Muse Spark Open, an 816B-parameter open-weight AI for code and media
Meta has released Muse Spark Open, an open-weight 816-billion-parameter multimodal model optimized for long-context coding and agentic tasks.
Meta and Apple collaborate on Muse AI models for Apple Silicon Macs
Meta announces Muse Spark Open, a multimodal AI model it calls the most capable US open-weight model to date.