Meta Releases Advanced AI Model That Runs Locally on Apple Macs

2026-10-16 companies

Menlo Park, Thursday, 15 October 2026.
Meta’s new 816-billion-parameter open-weight AI model, Muse Spark Open, can run locally on two connected Apple Mac Studio desktops through a rare technical partnership between the tech giants.

Meta Unveils Muse Spark Open

Meta Platforms Inc. (META) officially launched Muse Spark Open on October 15, 2026, delivering an 816-billion parameter natively multimodal open-weight AI model [1][2]. Built to support multi-agent orchestration and visual chain-of-thought processing, the release represents a strategic attempt by Meta to anchor enterprise software workflows around its open-weight AI ecosystem rather than proprietary walled gardens [1][5]. The model is described by the company as the most capable US open-weight model to date, optimized for complex reasoning and software development [2][3].

Collaboration with Apple

In a rare technical partnership, Apple and Meta announced a collaboration to enable the execution of this AI model on Apple high-end desktop hardware, specifically Mac Studio computers [6]. Apple and Meta have been collaborating for several months to enable this integration, a process facilitated by former Apple executive Daniel Gross [1]. This partnership marks a shift from their historically adversarial relationship, notably regarding privacy policy disagreements dating back to at least April 26, 2021 [1].

Technical Architecture and Hardware Integration

The Muse Spark Open model features 816B total parameters with 42B active parameters and a 1M-token context window [3][4]. It is a natively multimodal model capable of processing text, image, video, and audio inputs [3][5]. Utilizing the M5 Ultra’s 1.2 TB/s memory bandwidth and up to 512 GB of unified memory per Mac Studio, two linked machines can support the model, effectively functioning as a single supercomputer with 1 TB of memory [1].

Mac Studio Optimization

Optimization for Apple hardware allows Muse Spark Open to run on Mac Studio systems powered by the M5 Ultra chip, leveraging Apple latest MLX open-source array framework and unified memory architecture [1]. The model ships with Helion GPU kernels for SGLang, providing 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs [5]. Deployment requires significant hardware, specifically GPUs with at least 700 GB of VRAM for NVFP4 or 2 TB for BF16, though it is optimized for two Mac Studios with M5 Ultra using MLX [4].

Licensing Framework and Commercial Access

Meta released the Muse Spark Open model on October 15, 2026, which is licensed for use, modification, distribution, and commercialization under specific conditions [4]. The model is available immediately via API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI [5]. It is also available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models [5].

Commercial Terms

Licensees providing Model as a Service via API or similar interface access must enter a separate written agreement with Meta if their aggregate gross revenue exceeds 20 million USD during any consecutive 12-month period [4]. Licensees must prominently display Built with Muse Spark Open in user interfaces, documentation, or settings menus for any products or services utilizing the software [4]. Meta explicitly disclaims ownership of any outputs generated by the Muse Spark Open software [4].

Strategic Positioning Against Proprietary Models

This release follows the October 12, 2026 launch of two new proprietary AI models by Meta Superintelligence Labs, Muse Wonder and Muse Wonder Full [8]. Unlike the open-weight Muse Spark Open, the Muse Wonder model is a proprietary, hosted model not available for public download or self-hosting [8]. API pricing for the standard Muse Wonder is $4.00 per million input tokens and $20.00 per million output tokens, whereas Muse Spark Open on Baseten is priced at $1.25 per million input tokens and $4.25 per million output tokens [7][8].

Market Context

The total cost per million tokens for Muse Spark Open on Baseten combines input and output pricing, calculated as 5.5 [7]. Meta AI chief Alexandr Wang noted that while most open-weight models are too large for personal computers, Apple Mac Studios provide sufficient memory to run Muse Spark Open effectively [6]. Users are advised to implement system-level guardrails, as Meta recommends deploying Muse Spark Open as part of an AI system with guardrails tailored to intended use cases [4].

Sources

Meta Platforms Artificial Intelligence