I can't think of a better way to get Meta back open source than to work with Apple.
In the early hours of Friday, Meta suddenly released the Muse Spark Open, dubbed "the strongest open source major model in the United States". It's a native multi-modal model with parameter 816B, activated parameter 42B, 1M, specifically designed for code and intelligent object missions, and is good at long-term tasks and complex multi-step reasoning.
Meta hopes that in the future everyone will have the freedom to run advanced models at home. Zuckerberg said that the open-source version of Muse Spark was specifically optimized for Mac devices and that he worked with Apple's new CEO, John Ternus.
Specifically, Muse Spark Open is a model similar to Muse Spark 1.3, which uses the latest version of Apple's MLX framework to optimize the unified memory architecture of Apple's chips so that the two Apple M5 Ultra Mac Studio (512G) can run locally.
This is the first time I've ever heard of Hugging Face.
Apple's unified memory architecture allows the integrated GPU of the M5 Ultra to use up to 512GB of memory per Mac Studio, so the two machines can accommodate a complete large model. The chip's 1.2TB/s of memory bandwidth is enough to read the small bits of memory required for each token. Ultimately, the two Mac Studios run at the same speed as a supercomputer with 1TB of unified memory.
However, it is worth mentioning that the price of the two 512G Mac Studios has already exceeded $40,000.
In an interview, Meta AI CEO Alexandr Wang said they were "delighted to see parents back as good as they were at first". Wang explained that Apple and Meta AI had been working together for months. He also thanked former Apple executive Daniel Gross in particular for his role in promoting friendly cooperation between the two companies.
Meta has also helped developers adapt mainstream logic engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, and optimized for AMD, Apple, Intel, and Intel chips. Developers can easily customize further models using the PyTorch framework.
Key benchmark parameters:
It looks like Meta has a long way to go, and Muse Spark Open is not the strongest open-weight model in the world, with the programming, reasoning and other capabilities of Kimi K3, MiMo-V2.6-Pro, Qwen3.8 Flash Next and GLM 5.3 Flash.
It is worth mentioning that the Agent rating capability. As you can see from the official model card, the Muse Spark Open is even ahead of the Integrated Intelligence Score in some of the Agent indicators.
For example, Muse Spark Open uses Muse Code, Kimi K3 uses Kimi Code, Qwen 3.8 uses Claude Code. The actual capabilities of the Agent depend on the model, Harness, tools, context management and execution environment.
The value of a future open source model is not only based on the base model rankings, but also on whether it can enter the real Agent workflow.
References are given below:
© THE