Meta just unveiled the most powerful open-source weighted model in the United States!

  It's the first time in the history of the world.

It's been a while since we've had a big announcement from Meta:Muse Spark Open is officially open source.

The official positioning of the channel is extremely clear:The most powerful open-source weight model in the United States to date」。

The number of parameters is 816B (816 billion).

It's a million tokens.

Multimodal natives: natives support text, images, video and audio.

Designed for coding tasks and situations requiring complex reasoning, it is suitable for long-term tasks, complex software engineering problems and multi-step reasoning scenarios.

And the most surprising thing is,Two Apple Mac Studios can run this 816B super giant.。

You don't need machine rooms, you don't need expensive cloud computing, you don't need to pay API taxes.

In today's open-source environment, what is missing is a "super-intelligent body" that can do its job.

Muse Spark Open was born for this.

It focuses on long-range tasks, complex software engineering, and multi-step reasoning. It is customized for both coding and intelligent work.

Its core hardness indicators are as follows:

Total parameter 816B and activated parameter 42B。A typical MoE architecture. This makes it compatible with extremely high IQ and amazing reasoning efficiency.

Context window on 1M Token。One million in context, enough to fit into a whole set of large software source code, or dozens of heavy technical manuals. It remembers every instruction and never forgets.

Polymorphism of origin。It integrates text, images, audio, and video from the ground up. It understands UI design patterns, understands your voice needs, and reviews video content one at a time.

It's a six-sided warrior without a shortcut.

Additionally, the weight of the Meta model can be downloaded to Hugging Face and accessed from AWS SageMaker, Google Cloud, and Microsoft Foundry.

The mainstream API platforms such as Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together AI will also be online today.

What's more, it's a license plate.Muse Model License。

The Muse Model License grants the right to operate, modify, fine-tune, distill and redistribute this software, subject to its terms and conditions and use policies.

Its revenue terms are addressed to enterprises that provide models as services. When the license user and its affiliates have a combined total revenue of more than $20 million in any consecutive 12-month period, such as to provide Spark Open or its derivatives in the model service business, the license is required to notify Meta prior to delivery and to otherwise sign a written agreement.

This means:The vast majority of developers can use it for commercial use for free.Meta is delivering weapons directly to developers around the world.

The 816B model, in the past, has been a nightmare for ordinary developers.

You need to buy a lot of GPUs and build complex distributed clusters.

But this time, Meta and Apple teamed up to break the laws of physics.

This is the first time I've ever heard of the Muse Spark Open.You only need two Apple Mac Studios with M5 Ultra chips.

The secret lies in Apple's "unified memory architecture".

The M5 Ultra chip allows each Mac Studio to have up to 512GB of unified memory.

1TB of unified memory, plus M5 Ultra's up to 1.2TB/s of memory bandwidth.

In addition to the latest Apple MLX framework, the two are placed in a silver box on your desktop, directly embodied in a TB-level memory supercomputer.

Alexander Wang, the head of Meta AI, said:

Many powerful open-source models are too big for ordinary computers to run at all.

However, Apple's Mac Studio provides enough memory to make running Muse Spark Open very practical.

Of course, it's not cheap: one full-size M5 Ultra Mac Studio is priced at around $10,799, and two are over $20,000.

But for more than $20,000, buying a permanent online, unconnected, data-safe, top-of-the-line super-AI employee is obviously cost-effective for startups, independent developers, and video studios.

In a cluster of four Mac Studios, Muse Spark Open can generate more than 36 tokens per second, much faster than the average human reading speed.

In addition to Apple, it is fully compatible with other hardware, all perfectly suited to mainstream reasoning engines:

It's the first time I've seen a chip from NVIDIA, AMD, Intel.

Such as SGLang, vLLM, TensorRT-LLM, llama.cpp etc., especially in SGLang, it can even bring up to 3.95 times the core level of speeds for NVIDIA and AMD GPUs.

Muse Spark Open uses speculative decoding on four NVIDIA Blackwell GPUs to deliver more than 50,000 tokens per second to a single GPU and more than 600 tokens to a single user.

Super-smart, officially off the table!

Meta gives two extremely hardcore native, end-to-end intelligent workflows using Muse Code powered by Muse Spark Open:

Delivered real iOS modified chips to the open-source Mastodon client Ice Cubes to fix bugs, add features, and complete build;

The DaVinci Resolve is driven side-by-side, and the model examines the material step-by-step to determine the most productive moments and cuts the screens.

This means that you are no longer the user of the tool, but the mastermind of the intelligent body. The extreme privacy, the extreme power of performance is the ultimate form of End-Agent!

Behind the glowing technical indicators of the Muse Spark Open is a very dramatic business darkness.

In order for the model to run smoothly on the Mac, Meta and Apple, the centuries-old rivals of Silicon Valley, have shaken hands.

This alliance is rare and incredible.

It's important to note that apples and meta once did not tolerate water fires, and needles are relative.

Cook, the former CEO of Apple, has repeatedly accused Meta of violating user privacy in public.

Zuckerberg, meanwhile, has spoken out against Apple, ironically saying that "nothing good has been invented since the iPhone".

In 2021, Apple introduced strict privacy tracking restrictions (ATT) that directly hit Meta's advertising business. Meta estimated that the policy would result in a total loss of $10 billion in revenue in 2022.

But now the times have changed.

Today, Apple has welcomed its new CEO, hardware maniac John Ternus, while Meta's AI power has been given to young Alexander Wang.

In the face of the AI flood, old-fashioned complaints have been replaced by realistic interests.

For Apple, they missed out on a pioneering cloud model. Previous collaborations with Siri and OpenAI failed to live up to expectations. Apple urgently needed to prove to the world that the Mac is still the most powerful AI productivity tool on the map.

For Meta, Zuckerberg needed to break the "cloud" monopoly. If the AI model could only run in the cloud, the initiative would be forever in the hands of Amazon, Microsoft, and Google, who have huge machine rooms. Meta had to find a hardware terminal capable of supporting its supermodels.

So, one shot is a match.

Zuckerberg posted a photo of the Muse smartphone running on the Mac and sent an email directly to Apple's new CEO: "Thanks to the collaboration with John Ternus and Apple, the Muse Spark Open works great on the new Mac Studio!"

John Ternus responded quickly: "Congratulations to the team! It's amazing to be able to run a model of scale and scale like this at such an efficient level on the desktop.

When the two giants set their sights on the same goal, the whole industry will be shocked.

In 1984, Apple used the first Macintosh to free a computer from a large machine room and put it on a personal desk.

Today, in 2026, Meta and Apple are joining forces with Muse Spark Open to release super-AI from a huge cloud server cluster and put it back on the desktop.

The era of desktop-level superintelligence officially kicks off today. If you're a developer, entrepreneur or creator, you can run both Macs.

It's your personal supercenter, it's here!