Meta has released Muse Spark Open, an open-weight AI model designed for coding, multimodal reasoning and agentic work, with a launch message built around Apple’s high-end Mac Studio.
Meta announced Muse Spark Open on October 15, calling it the most capable U.S. open-weight model to date. The model has 816 billion total parameters, 42 billion active parameters and a 1 million-token context window, according to Meta and the model card published on Hugging Face.
It is a mixture-of-experts model built for long-running coding and agent workflows. The model card says Muse Spark Open accepts text, image, video and audio input, and returns text output. Meta is positioning it for tasks such as software engineering, tool use, long-horizon reasoning, multimodal analysis and synthetic data generation.
Model name: Muse Spark Open
Release date: October 15, 2026
Parameters: 816B total, 42B active
Context length: 1M tokens at inference
Inputs: Text, image, video and audio
Output: Text
License: Muse Model License
The weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta also says the model is available through hosted API providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
The Mac claim comes with a high hardware bar
The Mac angle is the attention-grabber. Meta says Muse Spark Open is the first model in its class to run on a desktop using two Mac Studios with Apple’s M5 Ultra chip and the latest MLX framework.
That does not make it a casual download for a MacBook. The Hugging Face model card says running the model on a user’s own GPU system requires at least 700GB of VRAM for the NVFP4 version or 2TB for BF16. The Mac path relies on Apple’s unified memory architecture instead of conventional discrete GPU memory.
Apple’s August Mac Studio announcement says the M5 Ultra version can be configured with 512GB of unified memory and 1.2TB/s of memory bandwidth. Meta says two such machines can hold the full model and behave like one larger local system through MLX.
Axios reported that Apple and Meta worked together for several months on the Mac support, a rare cooperation between two companies that have often clashed over privacy and computing platforms. For Apple, the demo gives Mac Studio a clearer role in local AI. For Meta, it gives the open-weight release a desktop story without relying only on Nvidia-style server setups.
Open-weight, but not unrestricted
Muse Spark Open is an open-weight release, not an unrestricted public-domain style release. Developers can access the model weights, but use is governed by Meta’s Muse Model License and usage policy.
The license allows many forms of commercial use, including copying, modifying, fine-tuning and creating derivative works, subject to conditions. One of those conditions requires products or services using the model to display Built with Muse Spark Open in the user interface, documentation, website or settings menu.
The license also creates a separate requirement for some Model-as-a-Service businesses. If a company or its affiliates offer model access through an API or similar service and pass $20 million in aggregate gross revenue during any consecutive 12-month period, the license says they must notify Meta and enter a separate written agreement before offering Muse Spark Open or derivatives as part of that business. The model card says this condition does not apply to internal use, research or commercial activity that is not Model-as-a-Service.
Meta’s usage policy also bars several categories of misuse, including fraud, spam, impersonation, malicious code and representing AI outputs as human generated. That is especially relevant for businesses planning to use the model in publishing or customer-facing systems.
What changes for local AI
The practical takeaway is not that every business can now run a top-tier model under a desk. The hardware requirements are too high for that. The change is that local AI is becoming more realistic for organizations with workstation budgets, sensitive workflows or a strong reason to avoid routing every request through a hosted model provider.
For most teams, hosted API access will still be the simpler path. It avoids the upfront hardware cost, setup work and maintenance burden. Local deployment becomes more interesting when the workload justifies the machine, when data control is a major requirement, or when developers need to customize and test model behavior without depending entirely on a third-party service.
This release also shows how AI hardware decisions are shifting. Apple’s pitch centers on large unified memory, while much of the AI infrastructure market has been built around discrete GPU clusters. Muse Spark Open does not settle that debate, but it gives developers another serious configuration to test.
The caveats are not small
The model card includes several warnings that should not be skipped. Meta says Muse Spark Open has additional safety post-training, but still recommends deploying it as part of a broader AI system with use-case-specific guardrails. For agentic workflows, Meta recommends safeguards such as human confirmation before irreversible actions.
The model card also says the model may produce inaccurate, biased or objectionable responses, and that it can still make errors in multi-step reasoning. Meta adds that downstream modifications could weaken safety behavior, so results from the released checkpoint should not be assumed to apply to fine-tuned or modified versions.
Training data is another area to watch. The model card says training data includes text, images, audio and video from public sources, internet-crawled material, third-party data, data derived from Meta’s products and services, and synthetic data curated or enriched by Meta personnel and external vendor networks.
For publishing workflows, that caution overlaps with Tech Help Canada’s guidance on AI-generated content and search quality: usefulness, verification and original expertise matter more than whether a draft came from a model.
What comes next
Meta has published benchmark results comparing Muse Spark Open with other open-weight models, but independent testing will decide how those claims hold up in normal developer workflows. The most useful tests will measure more than raw scores. They will need to cover setup complexity, inference speed, cost, stability, licensing friction and safety behavior after customization.
Get new small business insights by email
Practical ideas and useful articles to help you make better business decisions.
Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us
