Complete AI Training
Sign inGet my AI kit

Your job's AI kit

Get your AI kit

Tell us who you are and what you do. We show you your kit right away and email you the link: skills, prompts, AI agents, MCP servers and courses for your job.

500+ jobs ready, and we make a kit for any other job. No payment needed to look.

Share

AI news ·

Muse releases Muse Spark Open, an 816B-parameter open-weight model that runs on desktop Mac Studios

Meta released an 816-billion-parameter model that runs locally on two Apple M5 Ultra Mac Studios. Developers can now handle complex coding and video editing tasks entirely on-device, removing the need for cloud compute.

2 min read
CreativesIT and developmentProduct developmentWriters

Meta released the weights for Muse Spark Open today, putting an 816-billion-parameter multimodal model with a 1-million-token context window onto desktops. Two Apple M5 Ultra Mac Studios running the MLX framework can now run the model locally, a first for a model of this class. The release targets developers and creatives working on long-horizon coding tasks, complex software engineering, and multi-step reasoning workflows.

The model uses a mixture-of-experts architecture with 42 billion active parameters. It is available now through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

Running on Apple silicon

Apple's unified memory architecture allows the M5 Ultra's integrated GPU to use up to 512 GB per Mac Studio. Two machines linked together can hold the entire model, and the chip's 1.2 TB/s memory bandwidth keeps token generation fast. The result is a terabyte of unified memory behaving like a single desktop supercomputer.

Meta tested agentic workflows running entirely on-device. In one demo, Muse Code - powered by Muse Spark Open - shipped real iOS changes to Ice Cubes, an open-source Mastodon client, handling bug fixes, new features, and a full build. In another, the model drove DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and crop them vertically.

Inference and customization

The model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon. Developers can customize it using PyTorch and TorchTitan. The open-weight release ships with custom Helion GPU kernels in SGLang, delivering 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350.

"We're sharing this work because we want everyone to have access to the best models," the company said in its announcement.

Safety and licensing

Muse Spark Open received additional safety post-training specific to open-weight release. Meta said the model met the thresholds required by its internal scaling framework for release. The weights are available under the Muse Model License with limited commercial terms, which allows many developers to use the model for free.

Why this matters for developers and creatives

This release removes the cloud compute barrier for running a frontier-scale model. A creative team editing video or a developer shipping code can now run the same model locally that would otherwise require API credits. For anyone building agentic workflows - where latency and data privacy matter - two Mac Studios become a self-contained inference cluster. The practical ceiling for on-device AI work just moved up considerably.

Share