Meta released Muse Spark Open on October 15, 2026, and called it the most capable open-weight model in the United States. That is Meta’s own description. What makes the launch unusual is less the claim than the hardware story attached to it: the weights are public, but running them yourself takes roughly a terabyte of memory.
This post separates what is confirmed from what is vendor-reported, because the gap between the two is large right now.
What Muse Spark Open Is
According to TechCrunch’s launch report, the model has 816 billion parameters and a one million token context window. It is natively multimodal and aimed at complex coding and agentic work. Let’s Data Science adds that it is a mixture-of-experts design with 42 billion parameters active per token, which is why a model this large can still produce tokens at a usable speed.
Engadget rounds the size to 800 billion, so expect the figure to vary slightly between outlets. The weights are available on Hugging Face and through hosted inference partners, so you do not need your own machines to try it.
If you have followed Meta’s earlier release, note that this is a different thing from the Muse agent that launched in Meta’s own app. That one is a product you use. This one is a set of weights you can download.
The Hardware Problem
The headline number is memory. Engadget reports that running it locally needs about 1TB of RAM. Meta optimized the model for Apple Silicon, specifically the M5 Ultra Mac Studio, where the integrated GPU can address up to 512GB. Two machines together hold the full model.
The reported setups look like this:
- Two M5 Ultra Mac Studios, 512GB each: the configuration Meta describes, linked so they behave like one machine.
- Four M5 Ultra Mac Studios, 256GB each: an alternative, estimated by Engadget at around $38,000.
- Quantized builds: third-party reports mention lower-bit versions running on a single 512GB machine, at a cost to quality that has not been measured.
One detail matters for planning. Apple has not announced pricing or a ship date for the 512GB Mac Studio, only that it is expected in late October. Anyone budgeting for this today is budgeting against a price that does not exist yet.
Speed on a Home Cluster
Let’s Data Science reports Apple’s own measurements at an 8,192-token context. Two Macs in a pipeline setup managed about 15.7 tokens per second, two Macs with tensor parallelism about 25 tokens per second, and four Macs about 36.7 tokens per second. Those are usable for a single person working through a coding task. They are not a replacement for a hosted service handling many users at once.
These figures come from Apple’s test conditions, not an independent lab. Longer contexts will be slower, and the one million token window is a ceiling, not a promise of speed.
Open Weights Do Not Mean Unrestricted
The weights ship under the Muse Model License, which Meta describes as carrying limited commercial terms. One outlet reports that companies with more than $20 million in annual revenue that offer the model through an API need a separate agreement with Meta. I could not find the license text, and the Hugging Face page requires a login to accept the terms, so check the real document before building a product on it.
This is the same pattern worth watching across the industry. As we covered in our piece on gated AI models, access to the strongest systems is increasingly shaped by terms and conditions, not just by whether you can download a file.
What Is Still Unverified
Meta says its own benchmarks show the model beating the Chinese open models GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. Engadget notes these results have not been independently verified, and TechCrunch cites no specific scores. Meta’s example uses, such as bug fixes on a Mastodon client and footage review in DaVinci Resolve, are also vendor-supplied.
For most people the practical advice is simple:
- Try it through a hosted partner first if you only want to judge quality.
- Wait for independent benchmarks before treating “most capable” as settled.
- Read the license before any commercial use.
- Only consider local hardware if privacy or control justifies a five-figure outlay.
If cost is your main concern, a cheap hosted model may serve you better. We looked at that tradeoff in our Claude Haiku 5.5 breakdown.
Why It Matters
Muse Spark Open shows where open models are heading: bigger, more capable, and harder to run at home. The weights being public gives researchers and companies transparency and flexibility that closed models do not offer. Whether the quality holds up outside Meta’s own tests is the open question, and the next few weeks of independent results should answer it.
Sharing Is Caring