Half back to focus on open-weight models with Muse Spark Opena new artificial intelligence designed for programming tasks, complex reasoning and automation through agents.
The model, which It has 816 billion parameters in total, of which 42 billion are activei, has also been optimized to run locally on systems Apple Mac Studio with M5 Ultra chip. To run it in the reference configuration However, you need two machines equipped with 512 GB of unified memory eachfor a total of one terabyte.
The announcement confirms Meta’s desire to continue investing in the distribution of models with weights accessible to developers, in a market in which US companies compete with solutions developed by Chinese laboratories. The company calls Muse Spark Open “the most capable US open-weight model ever made“.
Muse Spark Open is a native multimodal model with a million-token context window. Its architecture includes 816 billion parameters in total, but activates 42 billion of them during processing. This configuration allows you to distinguish the total size of the model from the quantity of parameters involved in the individual inference steps, while requiring sufficient memory to accommodate the weights and other components necessary for execution.
Meta designed the system especially for programming, software engineering and workflows entrusted to AI agentsi.e. systems capable of planning and executing sequences of operations rather than simply responding to a single request. The goal is to tackle multi-step tasks, maintain context during extended sessions, and use external tools to complete multi-step tasks.
Click to enlarge
The company also cites multimodal reasoning and visual content processing capabilities. In benchmarks published by Meta, Muse Spark Open scored outperformed GLM 5.3 and Kimi K3 in some assessments related to agentic tasks and visual reasoning. Tests also include comparisons with other models, including Inkling, Nemotron 3 Ultra, Beam, Mistral 4 Large and Qwen3.8-2.4T-A95B.
One of the most relevant aspects of the launch is Optimization for Apple Silicon. Meta has collaborated with Apple to allow Muse Spark Open to run on Mac Studio equipped with M5 Ultra, taking advantage of the MLX framework developed for calculation on Apple hardware.
The reference system features two Mac Studios with 512 GB of unified memory each. Apple’s chip architecture allows the integrated GPU to use memory shared with the rest of the system, avoiding the rigid separation between system memory and VRAM typical of many discrete CPU- and GPU-based configurations.
According to Meta, each Mac Studio M5 Ultra has 1.2 TB/s of memory bandwidth. The combination of two systems allows the model to be distributed across a total capacity of one terabyte of unified memory, making it possible to run a model of this size locally without necessarily resorting to cloud infrastructure.
The configuration has a high cost. Apple has not yet communicated the precise price and availability date of the Mac Studio with 512 GB of memory, but a single 256 GB model costs 12,529 euros.
@finkd Congrats to the team! Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!
— John Ternus (@johnternus) October 15, 2026
Muse Spark Open weights are available at Hugging Face and through Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. The model can also be used via the Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI API platforms.
On the software front, Meta reports compatibility with several inference engines, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Support covers optimizations for AMD, Apple, Intel, and NVIDIA hardware, while developers can further customize the model using PyTorch and TorchTitan.
The open-weight distribution also includes custom, auto-tunable Helion GPU kernels for SGLang.
