Meta has released a new open-weight AI model called Muse Spark Open and says it can run locally on Apple Macs. The catch is the hardware. The full model needs about a terabyte of memory, which means a pair of top-end Mac Studios rather than an ordinary laptop. Here is what Meta announced on 15 October 2026.
What Is Muse Spark Open?
Muse Spark Open comes from Meta Superintelligence Labs. Meta calls it “the most capable US open-weight model to date.” Key facts from Meta’s announcement:
- Size: 816 billion total parameters, with 42 billion active at a time.
- Context window: 1 million tokens, which is the amount of text it can consider at once.
- Type: Natively multimodal, so it can handle more than text.
- Built for: Coding, long tasks, software engineering and multi-step reasoning.
What “Open-Weight” Means
Open-weight means Meta has published the trained model files, called weights, so developers can download and run them. It does not mean fully open source. Meta releases the weights under the Muse Model License, which has limited commercial terms. Meta says the terms let many developers use it for free. Read the licence before using it in a product.
How It Runs on Macs
Meta says it optimised the model for Apple’s MLX framework on Mac Studio computers with the M5 Ultra chip. A report says Meta and Apple worked together on the launch.
- Each Mac Studio can use up to 512 GB of unified memory.
- Two such Mac Studios, linked together, can hold the full model. Meta describes them as acting like one machine with around a terabyte of memory.
- Engadget reports that four 256 GB M5 Ultra Mac Studios can also run it, at roughly $38,000.
- Apple has said the 512 GB model arrives in “late October,” but no price has been announced.
So “runs on Macs” is true, but this is not something most people can do at home. The hardware cost puts it out of reach for individual users. Larger organisations that spend a lot on AI services may find it worthwhile if the model performs as Meta claims.
What Meta Showed It Doing
Meta paired the model with Muse Code, a coding terminal, and ran it on the device. Its demos included:
- Fixing bugs and adding features to an open-source Mastodon client for iOS.
- Reviewing and cropping video footage in DaVinci Resolve.
Performance Claims
Meta says its tests show Muse Spark Open beating other popular open-weight models. Engadget reports Meta claims it beats the Chinese models GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. These are Meta’s own benchmark results. Independent testing has not been reported yet.
Where to Get It
- Download weights: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
- Hosted APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
- Tools to run it: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth.
If you cannot afford the Mac hardware, the hosted APIs are the easier way to try it.
Safety
Meta says it added extra safety training because of the risks of releasing open weights. It says the model met the release thresholds in its internal safety framework. Detailed results are in the Hugging Face model card, and Meta says a full evaluation report will come later.
Frequently Asked Questions
What is Muse Spark Open?
An open-weight AI model from Meta with 816 billion parameters and a 1 million token context window.
Can I run it on my MacBook?
Not the full model. It needs about 1 TB of memory, which means two 512 GB Mac Studios or similar.
Is it free?
Meta says the licence lets many developers use it for free, but the licence has limits on commercial use.
Where can I download it?
The weights are on Hugging Face and several cloud platforms.
Are Meta’s benchmark claims proven?
They are Meta’s own results. Independent testing has not been reported yet.
Muse Spark Open shows that very large AI models can now run on desktop hardware, but for most people the practical option will be the hosted versions for now.
