Meta Releases Muse Spark Open Weights, Optimized for Apple M5 Ultra Mac Studio
On October 15, Meta released Muse Spark Open weights, calling it the strongest open weight model from the U.S. to date, optimized for the Apple M5 Ultra Mac Studio.
The model features a native multimodal mixture of experts architecture with a total of 816 billion parameters, approximately 42 billion activations per token, a context window of up to 1 million tokens, and supports text, image, audio, and video inputs with text output; it is designed for coding and long-cycle agent tasks, excelling in complex software engineering and multi-step reasoning.
In collaboration with Apple, the model can run on two M5 Ultra Mac Studios via the latest MLX framework, becoming the first of its kind to be deployable locally on desktop; the weights are available for download on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry, and are also accessible through API platforms like Baseten, Fireworks AI, and Groq.
The open version adds safety post-training to enhance robustness without system-level protection; developers can further fine-tune it using PyTorch TorchTitan, supporting various inference engines such as SGLang and vLLM.
Meta previously released a closed-source version of Muse Spark in April, followed by version 1.3 in August, and this open weight release fulfills its promise for subsequent open-source development; Zuckerberg emphasized in a public post that the model performs well on the new Mac Studio, thanking John Ternus and Apple for their collaboration.
This release is part of a product-driven expansion of local deployment capabilities, with Meta making large parameter model weights public and optimizing them for consumer-grade desktop hardware, shifting funding and attention from purely cloud-based APIs to locally running agents and coding tools; beneficiaries include developers wishing to deploy large models locally and the Apple hardware ecosystem, while competitors relying on closed cutting-edge models face pressure.
Source: Public Information
ABAB AI Insight
After the open-source Llama series, Meta shifted to closed-source cutting-edge models, with Muse Spark released in closed form in April 2026, followed by version 1.3 in August and a commitment to subsequent open weights; this Muse Spark Open release fulfills that path while collaborating with Apple's MLX to lower deployment barriers from data center GPUs to dual Mac Studios, continuing its emphasis on local operation in its "personal superintelligence" vision.
Resources are primarily directed towards safety post-training and hardware optimization, enabling the 816 billion parameter model to run on two M5 Ultras (each with 512GB unified memory) via Thunderbolt clustering; the motivation is to expand the practical usability of open weights, attracting developers to conduct coding and agent experiments locally rather than relying solely on cloud APIs, thereby re-establishing influence in the open-source community.
A similar path can be seen in the openness of large parameter MoE models like Mistral and Qwen, as well as Apple's previous local inference collaborations with other manufacturers on MLX; currently, Meta is in a phase of parallel development between closed-source cutting-edge models (like Muse Wonder) and openly available models, attempting to establish a desktop deployment advantage in the U.S. open weight arena through hardware optimization.
This essentially represents a technological substitution and industrial chain reconstruction: as large models shift from purely cloud-based services to locally operable ones, hardware unified memory and framework optimization become new breakthrough points; by opening weights and binding to Apple desktop clusters, Meta shifts some inference from data centers to the user side, altering the cost structure and control distribution of model deployment.
ABAB News · Cognitive Laws
The value of open weights depends on their ability to run locally
Hardware optimization is more decisive than parameter scale for adoption
Desktop clusters are replacing some data center inference