Meta releases 816-billion-parameter model – it already runs locally on Mac Studio
Meta this week released the weights for Muse Spark Open, a multimodal language model with over 800 billion parameters — and Apple's researchers have already shown it can run on a desk, not in a data centre. In tests published Thursday, four M5 Ultra Mac Studio machines generated more than 36 tokens per second at a context length of 8,192 tokens, according to Apple's own figures relayed via TechBooky and MacStories. The demonstration marks an unusually direct collaboration between Apple and Meta on local inference at scale — but every headline number comes from the companies themselves, and several central questions remain unanswered.
What was released, and by whom
According to MacStories (John Voorhees), Meta "has just released the weights for Muse Spark Open, a multimodal large language model with a mixture-of-experts architecture. It has over 800 billion parameters in total, with 40 billion active per token, and it can run on the newest Mac Studio machines" (MacStories).
Apple's contribution is MLX, Apple's open-source machine learning framework for Apple silicon. Apple's researchers say the model can run across a cluster of Mac Studio machines using MLX, according to TechBooky (TechBooky). The sourcing here is secondary: neither Meta's release announcement nor Apple's Machine Learning Research blog post is available directly, so all figures below are company claims relayed through these two accounts.
The scale problem in numbers
Muse Spark Open is too large to fit in a single ordinary workstation at native precision. Apple estimates the model would require more than 1.6 TB of memory for its parameters at native precision. Compressed to a four-bit format, the footprint drops to around 493 GB — and that is before counting the memory used during the actual generation of a response (Apple, via TechBooky).
This is where the mixture-of-experts (MoE) architecture becomes relevant. Although the model has 816 billion parameters in total, only a fraction is activated for each token. Apple states the figure is roughly 42 billion active parameters per token, according to TechBooky, while MacStories writes 40 billion active per token. The figures are not consistent across sources and are likely a matter of rounding — but they are reported here as stated, each with its source. Either way, the MoE principle is what makes the setup feasible: while the total weights must sit in memory, the compute per token is far lower than the total parameter count suggests.
How the cluster works
The key to splitting 493 GB of parameter data across multiple machines is a low-latency connection. According to TechBooky, MLX uses RDMA (remote direct memory access) over Thunderbolt 5, allowing several Mac Studio machines to share the inference workload. RDMA means the machines can read each other's memory directly, without an intervening software layer — precisely what is needed for parallelisation not to eat the gains.
Apple tested two parallelisation strategies: pipeline parallelism and tensor parallelism, and reported its peak speed with four M5 Ultra systems, TechBooky writes. In the published tests, the four machines generated more than 36 tokens per second at a context of 8,192 tokens. TechBooky itself notes that the figure measures output speed under a specified test configuration — it is not a general performance promise across all scenarios.
The results should also be read against the M5 Ultra chip itself. According to Apple's Machine Learning Research blog, cited by MacStories, "the M5 Ultra's time to first token surpasses the M3 Ultra by 4.6x, thanks to the M5 Ultra's Neural Accelerators". In multi-Studio setups, the M5 Ultra delivers up to 1.8x faster time to first token at very large contexts, and token generation is 1.6x faster than the M4 Ultra, according to the same sources.
What it costs — and what you actually need
This is where the demonstration meets reality. Running Muse Spark Open requires at least two machines, according to MacStories. Some of Apple's tests were conducted with the 512 GB variant of the M5 Ultra Mac Studio — which, according to MacStories, is scheduled to launch by the end of the year, with no price announced. For comparison, the 256 GB model with 2 TB of storage costs over $11,000. Quadrupling such a machine therefore lands far outside the consumer market, and it remains unclear what the unreleased 512 GB variant will cost.
What remains unanswered
The most striking figures — 816 billion parameters, 42 billion (or 40) active per token, 493 GB at four bits, and more than 36 tokens per second — are all Apple's own claims relayed through TechBooky and MacStories. No independent source has confirmed them. It is also unclear exactly when Meta's weight release and Apple's blog post occurred; TechBooky says only that the tests were "published Thursday".
Several practical questions are likewise unanswered in the sources: How good is the model's actual output, particularly after four-bit quantisation of an 816-billion-parameter network? How stable is the latency in real-world use, not just under the benchmark configuration? And how do pipeline and tensor parallelism differ in practice? TechBooky notes that the published measurement covers output speed under one specific test, so generalisation to other workloads is not yet documented.
What does remain is a concrete technical demonstration: a model on the border between research frontier and production systems can split its work across a few desk-side machines over a Thunderbolt cable. Whether this becomes a practical route to local inference at frontier scale, or an impressive benchmark without broader application, remains to be seen.