Meta shipped Muse Spark Open on October 15, 2026, with a big claim attached: “the most capable U.S. open-weight model to date,” in Meta’s own words (MacRumors).
The model carries 816 billion total parameters, 42 billion active per token.
And a 1 million-token context window, with weights downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden.
And Microsoft Foundry Models (Meta). The detail that changes your planning: Meta worked with Apple to optimize it for Mac hardware. And the documented floor is two Mac Studio machines with M5 Ultra chips and 512 GB of unified memory each.
What Shipped, and What “Open Weight” Buys You
Muse Spark Open is natively multimodal on the input side.
It accepts text, images, video, and audio, and returns text only (Heise). If your pipeline needs generated images or audio out, this is the wrong tool. And no benchmark table will fix that.
Day-one availability is unusually wide. Beyond the raw weights, the model is live on seven API platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI (Meta).
On the inference side it supports SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp. And Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon.
That last list is the quiet headline.
Four chip vendors, llama.cpp, Unsloth. And an Apple-specific path means Meta built this expecting other people to run it on hardware Meta does not control. That is a genuine commitment, not a press-release gesture. One caution though: “most capable” is Meta’s own description, not an independent ranking. Treat it as a vendor claim until you test it on your workload.
Two Mac Studios, One Desk-Sized Cluster
The Apple collaboration is the part worth your attention.
The Mac implementation runs through Apple’s MLX framework, which pulls the weights directly from Hugging Face.
So setup is closer to a package install than a cluster deployment (Meta).
Meta’s documented configurations are specific: two Mac Studio machines with M5 Ultra and 512 GB of unified memory each, or four Mac Studios with 256 GB each (Meta dev docs). MLX splits the model across the Macs using RDMA over Thunderbolt 5, which requires macOS 26.2 or later and a direct link between every pair of machines. So this is a cabled cluster on your desk, and with four machines, every pair needs its own link. Nobody is doing this over Wi-Fi.
Why it works at all comes down to memory. Apple’s unified memory architecture lets each M5 Ultra’s integrated GPU use up to 512 GB.
So two machines together expose a terabyte of unified memory. And the chip’s 1.2 TB/s of bandwidth reads the small active slice each token needs (Meta).
At 4-bit precision the model needs about 493 GB of memory. And Apple measured roughly 25 tokens per second on two machines and about 37 on four (Plotly).
My read on those numbers: 25 tokens per second is fine for one person prompting, summarizing, and coding against it.
It is not a serving tier for concurrent users. The four-machine setup buys you 50 percent more throughput and a worse cabling problem, which tells you this path is aimed at individuals and small teams, not infrastructure.
The Memory Math Comes Before the Purchase Order
Here is the honest framing for a small operator. A 493 GB footprint at 4-bit is why the minimum is two maxed-out machines. And that class of purchase only makes sense in specific situations. The first is data control: if your documents legally or contractually cannot leave your building, a desk cluster running open weights is now a real option where a year ago it was a fantasy.
The second is exit optionality, and this matters even if you never buy a single Mac Studio.
Seven API hosts carrying the same open weights changes the pricing dynamic.
When a provider knows you have a credible walk-away path, including running the model yourself, their price has a ceiling. Open weights are your negotiating position even when you never exercise them. That is the durable value of this release for a lean operation, more than the parameter count.
So the sequencing question answers itself.
You test through an API this week, you benchmark on your actual tasks. And hardware becomes a conversation only after recurring volume or data rules force it. Buying machines first inverts the decision and locks you into a bet on one model, in a market where the interesting open-weight release changes every few months.
What I’d Do This Week
Start with a real workload, not the vendor’s benchmarks.
Route one genuine task through OpenRouter or Together AI, something with your data and your edge cases. And judge the output directly. Test the 1 million-token context window claim with a document long enough to actually stress it. Because no ranking page has demonstrated that window under load and you should not take it on faith either.
Then write down what you’re paying per week at current usage.
Recheck that number in a quarter against the other hosts, since seven providers competing on the same weights is where your savings live. If your data rules ever harden, the two Mac Studio configuration is your fallback plan.
And you now know its exact requirements: M5 Ultra, 512 GB each, macOS 26.2 or later, Thunderbolt 5 between every pair (Meta dev docs).
The benchmark race in this week’s coverage is noise. Signal: that a major U.S. lab shipped a model engineered from the start to run in a spare room, on hardware a person can actually buy. And told you exactly how. Run the API test, log your costs. And let your own numbers decide whether a desk cluster ever enters the picture.
