Meta just released one of the most powerful open AI models ever built. But there’s a catch: “open” doesn’t mean you can actually run it in your living room.
On October 15, 2026, Meta unveiled Muse Spark Open, an 816-billion-parameter multimodal model that the company calls “the most capable US open-weight model to date”. It’s a direct shot at China’s best open models like Qwen, Kimi, and GLM, and it represents a major moment for open AI. But for anyone hoping to download it onto their home PC and start experimenting, the reality is far more complicated—and far more expensive.
What Is Muse Spark Open?
Muse Spark Open is a mixture-of-experts (MoE) transformer with 816 billion total parameters, of which roughly 42 billion are active per token. It handles text, images, video, and audio, and features a staggering 1 million token context window. Meta designed it specifically for coding and agentic workflows—long-horizon planning, tool use, and complex multi-step software tasks.
The weights are available on Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry, and the model can be accessed through APIs from partners like Fireworks AI, Groq, and Together AI. On benchmarks, it holds its own against the best Chinese open-weight models, particularly on agentic and knowledge tasks.
The Hardware Reality: Two Mac Studios Minimum
Here’s where the dream of running it at home falls apart. Meta’s own hardware guide calls for two Mac Studios with 512GB of unified memory each, or four Mac Studios with 256GB each, linked by Thunderbolt 5. The model needs about 493GB of memory in a 4-bit format, and even then, it cannot fit on a single Mac Studio—the extra memory required during inference pushes it over the edge.
To put that in perspective, a Mac Studio with an M5 Ultra chip and 256GB of RAM costs at least $10,799. A 512GB version costs more. And you need more than one machine. Apple’s own testing found that four Mac Studios generate just over 36 tokens per second with the model, while two machines running tensor parallelization manage about 25 tokens per second. That’s usable, but it’s not fast—and it’s not cheap.
For those hoping to use consumer GPUs, the situation is even bleaker. The full BF16 checkpoint weighs about 1.65TB, and Meta’s serving recipe requires 8 × B300 GPUs with roughly 2TB of GPU memory. Even the NVFP4 quantized version needs 4 × B300 GPUs and at least 700GB of GPU memory. As one analysis bluntly put it, “普通消费级显卡就别想全权重部署，就算量化压缩，内存门槛依然劝退绝大多数人”—ordinary consumer graphics cards shouldn’t even think about full-weight deployment, and even with quantization, the memory threshold will deter the vast majority of people.
Can You Run a Smaller Version?
Technically, yes—but with major compromises. There are GGUF quantized versions available, including a 2-bit quantization at around 309.9GB that requires 512GB of unified memory or more. Another version needs ≥370GB of RAM/VRAM for fast inference, though it can run via memory-mapped paging on a 192GB Apple M2 Ultra—just very slowly.
These aren’t “home computer” solutions. They’re workstation-class setups. The smallest local set still requires hundreds of gigabytes of memory, far beyond what any typical laptop or desktop has.
The “Open” Caveat: A Limited License
There’s another wrinkle. While the weights are open, the license isn’t fully open. Meta released Muse Spark Open under the Muse Model License, which includes “limited commercial terms”. Specifically, companies with more than $20 million in gross revenue face restrictions on commercial use. So even if you could run it, using it commercially at scale isn’t as simple as downloading a file and firing up a terminal.
What This Means for Ordinary People
The honest answer is: no, ordinary people cannot run Muse Spark Open at home. Not on a gaming PC. Not on a MacBook. Not on a single high-end workstation. The model is simply too large, and the memory requirements are too steep.
That doesn’t make the release meaningless. It’s a significant moment for the open-weight AI ecosystem, and it gives developers and researchers a powerful new tool to build on—through APIs, cloud platforms, or institutional hardware. Meta AI chief Alexandr Wang acknowledged the gap himself, noting that numerous high-performing open-weight models exceed what most personal computers can handle, but Mac Studios offer enough memory to make running Muse Spark Open “realistic”. “Realistic” here means a multi-machine setup that costs tens of thousands of dollars.
For the rest of us, the practical path is the cloud. Muse Spark Open is available through multiple API providers, and that’s how most individuals and small teams will interact with it—renting access rather than owning the hardware.
Muse Spark Open
is a technical achievement and a statement of intent from Meta. It proves that open-weight models can compete at the frontier. But it also proves something else: the gap between “open” and “accessible” is wider than ever. The most powerful open models now require infrastructure that rivals small data centers, and the era of running cutting-edge AI on a home computer may be slipping further out of reach.
