Meta wants to challenge China’s AI dominance with its new Muse Spark Open model

Meta has launched Muse Spark Open, an 816-billion-parameter open-weight AI model designed for advanced coding, complex reasoning and AI agent workflows. The company claims it is the most capable US open-weight AI model to date.

Meta has introduced a new open-weight artificial intelligence (AI) model for complex coding, multi-step reasoning and autonomous tasks, called Muse Spark Open. Launched on October 15, 2026, the model will serve as a way to reinforce Meta's standing in the open-weight AI industry, where Chinese AI developers like Alibaba's Moonshot AI and Z.ai have made progress. According to Meta, Muse Spark Open is the most powerful open-weight model ever released by a U.S. company.

The model has 816 billion total parameters, with 42 billion active parameters, and supports a context window of one million tokens. According to Meta’s official announcement, developers can download its model weights, customise its capabilities and run it on supported hardware. It's also been optimised for Apple's M5 Ultra-powered Mac Studio systems and lets you use the model on the local machine rather than all the time in the cloud.

Meta Muse Spark Open targets coding and AI agents

According to Meta, Muse Spark Open is designed to serve long jobs, complex software development, and multi-step reasoning. The capabilities are targeted at developers creating AI agents that can manage workflows with minimal human interaction.

According to the official blog post, the model is compatible with multiple inference engines, such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX and llama.cpp. The model can also be customised by the developer using the PyTorch framework.

Meta has added customisable, auto-tunable Helion GPU kernels to SGLang. The company claims that these will provide kernel-level performance boosts from 1.03x to 3.95x on Nvidia and AMD GPUs and a 3% end-to-end performance boost on AMD MI350.

Pragya Singha Roy