Meta's Giant AI That Needs 1 TB of RAM: 816 Billion Parameter Muse Spark Open Unveiled

A giant step that redraws the boundaries of open-weight large language models has been taken. Meta has officially announced Muse Spark Open, which it describes as the most capable open-weight AI model ever developed in the United States. Reaching 816 billion parameters in total, this new-generation model was built specifically for complex software engineering, multi-step reasoning and long-running autonomous tasks. The most talked-about aspect of the model, however, is that thanks to deep optimization done with Apple, it can run fully locally on two Mac Studio computers with M5 Ultra chips, using a total of 1 TB of unified memory.

For organizations and advanced developers who want to run a frontier model on their own hardware without depending on cloud servers, Muse Spark Open turns the idea of a desktop supercomputer into reality.

816 Billion Parameters and a 1 Million Token Context Window

Meta's new flagship, Muse Spark Open, combines the efficiency of the Mixture-of-Experts architecture with a huge knowledge capacity. The key figures in the model's technical architecture are:

• Total parameter count: 816 billion parameters
• Active parameter count: 42 billion active parameters engaged at each inference step
• Context window: A 1 million token capacity that can analyze huge codebases, documents of hundreds of pages or long videos in one go
• Native multimodal structure: The ability to process text, code and image/video data within the same architecture without needing external plugins

According to benchmark results shared by Meta, Muse Spark Open outperforms existing open-weight rivals in many tough categories while delivering performance quite close to closed-source commercial models. The model's edge is especially notable in using external tools (tool use), managing long-running agent workflows and designing software architecture from scratch.

A Local Supercomputer With 1 TB of Unified Memory From Two M5 Ultra Mac Studios

Running an 816 billion parameter model locally requires extremely expensive enterprise server racks with traditional graphics card (VRAM) architectures. But joint work between Meta and Apple engineers managed to bring this huge model down to desktop size.

Mac Studio workstations with Apple's new M5 Ultra chip let the integrated graphics unit (GPU) directly access up to 512 GB of memory per system, thanks to the unified memory architecture. Two M5 Ultra Mac Studios connected through Apple's MLX machine learning framework split the weights between the two machines and create a pool of 1 TB of unified memory in total.

In this setup, not only memory capacity but also data transfer speed plays a critical role. The 1.2 TB/s memory bandwidth offered by the M5 Ultra chip lets the billions of weight values needed for each generated token be read without delay. Meta stresses that these two compact desktop systems, once connected, behave like a single AI supercomputer. Although Apple has not yet announced the official price of the M5 Ultra Mac Studio configuration with 512 GB of memory, it clearly offers a much more accessible and quieter local alternative compared with enterprise GPU servers.

Helion Kernel Optimization for Nvidia and AMD Graphics Cards

Of course, Muse Spark Open is not limited to the Apple Silicon ecosystem. Meta added full support for all popular inference engines in the industry so that the model runs at peak efficiency in data centers and on different hardware setups. The model comes compatible from day one with SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth and Apple MLX.

In particular, the custom auto-tunable Helion GPU kernels offered within SGLang create a serious performance jump on Nvidia and AMD accelerators:

• Kernel-level speedups between 1.03x and 3.95x are achieved on Nvidia and AMD graphics cards.
• On enterprise-grade AMD Instinct MI350 accelerators, end-to-end inference performance improves by a net 3 percent.
• For teams that want to fine-tune the model on their own datasets, the PyTorch and TorchTitan toolkits are provided ready to use.

Real-World Scenarios With Muse Code and DaVinci Resolve

To show what its new model can do in real desktop workflows beyond theoretical tests, Meta also shared striking use cases. Running on Muse Code, the company's own AI coding terminal, Muse Spark Open was put directly to work on Ice Cubes, a popular open-source Mastodon client. The model scanned the project code, fixed existing bugs, coded new features and completed the app's build process from start to finish on its own.

Another impressive scenario outside coding was seen in an integration with the professional editing software DaVinci Resolve. Muse Spark Open analyzed raw video footage frame by frame, identified the most important moments in the video and automatically cropped and edited the footage into a vertical format for social media posts. Being able to complete all of this directly on a local computer without sending any data to the cloud is a huge advantage for studios and software companies with strong data privacy requirements.

How to Download and Use Muse Spark Open

For developers who want to try the model right away or integrate it into their apps without investing in hardware, cloud API support was rolled out at the same time. Muse Spark Open is currently available via API calls on OpenRouter, Together AI, Fireworks AI, Groq, Baseten, Modal and Nebius.

Those who want to download the model to their own infrastructure and customize it can get the weight files from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta is releasing its new model under a custom license called the Muse Model License; this license gives individual developers and researchers free usage rights, while it includes special terms for commercial use above a certain scale.