US-based Meta Releases 816-Billion-Parameter Open-Weight AI Model That Supports Local Running on Apple Computers

en.Wedoany.com Reported - On October 15, US technology company Meta officially released the open-weight artificial intelligence model Muse Spark Open. The model has a total parameter scale of 816 billion, 42 billion activated parameters, and supports a 1-million-token context window. The model is designed for complex software development, multi-step reasoning, and AI agent tasks, and has been optimized for the Apple M5 Ultra chip, enabling local deployment through two Mac Studio workstations.

Muse Spark Open adopts a native multimodal architecture and can perform long-process programming, software engineering, and autonomous task handling. Meta has integrated the model into its AI programming terminal Muse Code and completed multiple local running tests, including performing program bug fixes, new feature development, and complete software builds for the open-source Mastodon client Ice Cubes, as well as using the DaVinci Resolve video editing software to analyze footage frame by frame, screen clips, and complete vertical cropping.

In terms of hardware adaptation, the model runs using Apple's MLX machine learning framework. A single Mac Studio equipped with the M5 Ultra chip can be configured with up to 512GB of unified memory, and two devices combined provide approximately 1TB of memory capacity. The M5 Ultra chip's memory bandwidth reaches 1.2TB/s, which can support data reading requirements during model inference.

Meta also optimized inference for different chip architectures, and the model is compatible with inference frameworks such as SGLang, vLLM, TensorRT-LLM, and Apple MLX. On NVIDIA and AMD graphics processors, custom Helion GPU kernels achieve 1.03x to 3.95x kernel-level acceleration, among which end-to-end inference performance on the AMD MI350 platform improves by about 3%.

The Muse Spark Open model weights are now available for download through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, and have also been integrated into inference service platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The model uses the Muse Model License, allowing developers to use and customize it under limited commercial terms.