Meta releases 816B-parameter Muse Spark Open AI model for Apple Silicon

Meta’s new open-weight model supports multimodal tasks and runs locally on dual Mac Studios, aiming to compete with Chinese AI leaders.

Threat Times Systems DeskOctober 15, 20262 min read

Share LinkedIn Reddit

Illustration: A sleek silver computer tower sits on a clean wooden desk next to a matching desktop monitor. — AI-generated illustration: Threat Times

Key points

Muse Spark Open has 816 billion parameters and supports text, image, video, and audio inputs.
The model is optimized for Apple’s M5 Ultra hardware, requiring two Mac Studios with 512 GB of memory each.
Meta claims the model outperforms Qwen 3.8, GLM 5.3, and Kimi K3 on specific coding and agentic benchmarks.

Meta has released Muse Spark Open, an open-weight artificial intelligence model containing 816 billion parameters. The company describes the release as its most capable US open-weight model to date, designed specifically for coding and agentic tasks. This launch marks Meta’s return to the open-weight space following its previous Llama series releases.

The model processes text, images, video, and audio with context windows of up to one million tokens. According to Meta, it achieves benchmark scores that surpass leading Chinese open-weight models, including Qwen 3.8, GLM 5.3, and Kimi K3. Despite these gains, the company notes that Muse Spark Open still trails behind Meta’s proprietary, paid Muse Wonder Full models in overall capability.

Technical Specifications and Hardware Requirements

Muse Spark Open utilizes a mixture-of-experts architecture with 42 billion active parameters out of its total 816 billion. The model is optimized for execution on Apple’s M5 Ultra hardware using the MLX framework. Running the model locally requires significant computational resources, specifically two Mac Studios, each equipped with 512 GB of unified memory.

This hardware requirement aligns with earlier reporting by Threat Times on October 15, 2026, which noted that Meta’s largest US model required a terabyte of RAM for local execution. The current release specifies the dual-Mac Studio configuration as the method for local desktop execution, leveraging the collaborative optimization efforts between Meta and Apple announced on that same date.

Licensing and Commercial Usage

Meta has made the model weights available on Hugging Face under a license that allows free usage for most purposes. Users must include the label "Built with Muse Spark Open" when deploying the model. However, commercial restrictions apply to large-scale service providers.

Companies generating more than $20 million in annual revenue that wish to offer Muse Spark Open as an API service must negotiate a separate agreement with Meta. This tiered licensing structure aims to balance open accessibility with commercial control, distinguishing it from fully permissive open-source licenses.

What should you do?

Security and technology professionals should evaluate the model’s performance on specific coding and agentic tasks against their current infrastructure. Organizations considering local deployment must verify that their hardware meets the dual Mac Studio requirement with sufficient unified memory.

What to do and how to stay safe: Meta

Review the model’s benchmark scores on JobBench, MCP-Atlas, HLE-Diamond, and ZeroBench to assess suitability for specific coding tasks.
Verify that your organization’s hardware infrastructure includes two Mac Studios with 512 GB of unified memory each for local execution.
Consult legal counsel to determine if your annual revenue exceeds the $20 million threshold, which requires a separate commercial agreement with Meta.
Monitor the Hugging Face repository for updates to the model weights and any changes to the licensing terms or usage labels.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.