Meta's 816-Billion-Parameter Model Runs on Two Mac Studios: The Memory Arithmetic Behind 'Local'

Meta's Muse Spark Open, released on 15 October 2026, has 816 billion parameters but computes with only 42 billion of them for each token it generates. The second number is the one that earns the headlines about speed; the first is the one that sets the price of running the model at home. Every one of the 816 billion weights has to sit in fast memory, because the next token might be routed through any part of the network. At 4-bit precision, ForkLog reports, the parameters alone need about 493 GB, which is why Meta's showcase hardware is two Apple M5 Ultra Mac Studios with 512 GB each rather than one.

This piece does the arithmetic that the launch coverage mostly skipped: what 'local' requires in memory, how fast the reported setups run compared with what memory bandwidth would predict, what the hardware plausibly costs, and what the licence asks of businesses. A note on sourcing before the numbers. The specifications come from ForkLog (which relays an Engadget cost estimate), MacRumors and YourStory. None of the three reports a benchmark score for this model, and we found no independent evaluation. Where a figure below is our own calculation from those published numbers, it is labelled as such.

What Meta released, and what is only Meta's claim

According to MacRumors and ForkLog, Muse Spark Open is a mixture-of-experts model from Meta Superintelligence Labs with 816 billion total parameters, 42 billion active per token, a one-million-token context window, and input in text, images, audio and video. Meta positions it for coding and agentic work. The weights are on Hugging Face and on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted API access is offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI (MacRumors). YourStory adds that the model runs on inference engines including vLLM, SGLang, Apple MLX, TensorRT-LLM and llama.cpp, and ForkLog says Meta also optimised it for AMD, Intel and Nvidia chips.

Meta describes the model as 'the most capable US open-weight model to date', according to MacRumors and ForkLog; ProPakistani gives the wording as 'the most powerful U.S.-developed open-weight AI model currently available'. That is a company claim, and none of the articles we read backs it with a comparison table. ForkLog supplies the relevant context: three days earlier, on 12 October, Meta introduced Muse Wonder and Muse Wonder Full and claimed they beat GPT-6 Astra and Claude in several tests, while Artificial Analysis rated them below Claude Opus 5.5 and Abacus AI chief executive Bindu Reddy questioned Meta's results. Those are different models, so nothing follows directly about Muse Spark Open, but it is a reason to treat 'most capable' as unverified until independent testers publish.

YourStory also reports, in paraphrase, that Meta ran safety training and evaluation under its Superintelligence Scaling Framework and says the model met its release thresholds. That too is Meta's own statement of its own process.

Why 42 billion active parameters still means about 493 GB

A mixture-of-experts model splits much of its network into many specialist sub-networks. For each token, a small router picks a handful of them to do the work, and the rest sit idle for that token. That is where the 42 billion comes from: roughly 5.1 per cent of the 816 billion (our arithmetic) is exercised per token, so the compute cost per token resembles that of a much smaller dense model. But the router's choice changes from token to token and cannot be predicted far ahead, so every expert has to stay resident in memory. Fetching one from storage on demand would be far slower than reading it from unified memory. Compute scales with the active count; memory scales with the total.

The raw arithmetic at 4 bits per weight is 816 billion times half a byte, or 408 GB. ForkLog reports about 493 GB for the 4-bit version used in testing, which implies an effective 4.8 bits per weight (our calculation: 493 GB times 8, divided by 816 billion). The sources do not explain the difference; the usual reasons are quantisation scaling factors and some tensors kept at higher precision, but that is our inference, not a reported fact.

That 493 GB is a floor, not a requirement. The machine also needs memory for the operating system, the activations, and the key-value cache that grows with the length of the conversation. A single 512 GB Mac Studio would leave about 19 GB for everything else (512 minus 493), and ForkLog states the model does not fit on one machine once intermediate computation is counted. Two 512 GB machines give 1,024 GB, leaving roughly 531 GB of headroom; ForkLog also reports a four-device configuration of 256 GB machines, which totals the same 1,024 GB.

How fast it runs: what bandwidth predicts versus what was measured

Generating text is mostly limited by how quickly the weights needed for each token can be read out of memory. ForkLog gives each M5 Ultra Mac Studio a memory bandwidth of 1.2 TB/s. If the 42 billion active parameters are stored at the same effective 4.8 bits as the whole model, each token requires reading about 25 GB (42 billion times 4.83 bits, divided by 8). On a single machine with 1.2 TB/s, that reading alone would take about 21 milliseconds, a ceiling of roughly 47 tokens per second. This is a back-of-envelope figure from our arithmetic: it ignores the cache, router overhead and the link between machines, and assumes the active experts are read at the average precision.

Now compare it with what ForkLog reports, all at a context length of 8,192 tokens. Two Mac Studios using a pipeline scheme, where each machine holds half the layers and only one works at a time, reached about 15.7 tokens per second. Against the single-machine ceiling of about 47, that is roughly a third. Two machines using a tensor scheme, where both work on every layer at once over a Thunderbolt 5 link with RDMA, reached about 25 tokens per second. Their combined bandwidth of 2.4 TB/s would raise the ceiling to roughly 94, so 25 is about a quarter of it. Four devices reached about 36.7 tokens per second, though ForkLog does not name the scheme and the machines have 256 GB each, so we cannot compute a ceiling or call it like-for-like.

Two readings follow, and both come from the arithmetic rather than from any source. First, the measured speeds sit well below the bandwidth ceiling, which is what one expects when machines must exchange data every layer; it is a description of overhead, not a criticism of the demo. Second, adding hardware buys less than proportionally: going from two machines (tensor, 25 tokens per second) to four (36.7) is a 1.47-fold gain for twice the devices, with the caveat that the configurations differ. For one person, 25 tokens per second is several times faster than a human reads. For a team sharing one rig, or for an agent that sends very long prompts, the figures say little, because they describe a single stream of generation at 8,192 tokens, and no source reports how quickly the machines process a long prompt in the first place.

What it costs

The only price signal is an Engadget estimate, relayed by ForkLog, of about $38,000 for the four-device, 256 GB-each configuration. MacRumors, however, says the 256 GB Mac Studio already costs more than $10,000, and four machines at that price would exceed $40,000. The two statements do not line up exactly, which is a reason to treat '$38,000' as an order of magnitude rather than a quote. The two-machine, 512 GB configuration that Meta showcases cannot be priced at all: ForkLog says the 512 GB Mac Studio is not yet on sale, and Apple has promised a release by the end of October without naming a price.

A break-even against renting the model is therefore not computable from published figures, and we are not going to invent one. The hosted providers named above publish per-token prices, but none of the sources we read gives them for this model. The formula the reader can apply is simple: hardware cost divided by what the same tokens would cost through an API, at the reader's real monthly volume. As an upper bound on volume, one stream running flat out at 25 tokens per second for 24 hours is about 2.16 million tokens a day (25 times 86,400). Most desks are idle most of the day, which pushes the break-even later, not earlier.

The licence and the privacy trade-off

The weights are free to download under what Meta calls the Muse Model License. ForkLog reports that commercial use is permitted with conditions, and that a company with annual revenue above $20 million that offers the model to others through an API needs a separate agreement with Meta. YourStory advises businesses to read the licence before building products on it, and MacRumors describes the terms as limited commercial use. The sources we read do not state how the licence treats a large company using the model internally rather than selling access, so that question needs the licence text, not a news article.

On privacy, ProPakistani reports Meta AI chief Alexandr Wang saying that Meta has 'no control or visibility' into how the model is used on local Macs. Run on your own hardware, the model does not send prompts to Meta. That is the strongest argument for running locally: confidentiality that does not depend on a vendor's retention policy.

The counter-argument is in the distribution list. A model that needs roughly $38,000 of Apple hardware to run privately will, for most organisations, be used through one of seven API providers or one of four cloud model catalogues, where the privacy position is whatever that host promises. In that case 'open' mostly means a choice of vendors and an exit option, not a machine under the desk. Both things can be true: the local option matters for the few who need it, and the hosted option is how most of the usage will arrive.

What the evidence does not establish

• Quality: no source reports a benchmark for Muse Spark Open, and none says how much the 4-bit version loses against full precision.
• Long context: speeds were measured at 8,192 tokens, not the advertised one million, and prompt-processing speed is not reported.
• Cost: the four-device estimate is secondhand and inconsistent with MacRumors' per-machine figure; the two-machine setup has no price because the 512 GB machine is not yet on sale.
• Leadership claim: 'most capable US open-weight model' is Meta's wording and has not been independently tested in anything we read.
• Licence scope: the $20 million clause is reported for API providers; internal-use terms are not described in the coverage.

The practical takeaway is narrower than the launch framing. Running an 816-billion-parameter model on a desk is real, and the memory arithmetic shows why it takes two top-end machines rather than one. Whether the result is worth the cost depends on measurements nobody has yet published: quality, long-context behaviour and the price of the hardware itself. Watch for the 512 GB Mac Studio's price at the end of October and for independent benchmarks; those two things will settle most of what this analysis has to leave open.