Hardware · verified 2026-10-17
What running Muse Spark Open actually costs
The launch post says two M5 Ultra Mac Studios can run Muse Spark Open. That sentence is true and incomplete: it is true in the sense that two machines can hold the weights, and incomplete in the sense that the useful configuration Apple measured is four. This page prices every route from a hosted endpoint to a 128 GB laptop, and says who measured what.
Meta cookbook, Apple ML Research and two repository authors Read 2026-10-17
Why this model is different
Most open-weight releases are a download and a quantisation away from a laptop. This one is not, for a structural reason: it is a sparse mixture of experts with 816B total parameters. Only 42B are active per token, so the arithmetic per token is manageable — but every expert has to be stored somewhere before the router can choose it. Compute tracks 42B; memory tracks 816B.
That is why the same model produces five very different hardware stories. A 4-bit format gets the weights to about half a terabyte, which is two top-end Macs. A 1-bit quant gets them under 200 GB, which is two small NVIDIA nodes. And a laptop can run it at all only by keeping three quarters of the experts on disk and streaming them in.
The five routes, priced
| Route | Hardware | Memory | Measured speed | Verdict |
|---|---|---|---|---|
| Hosted API | None | None | Provider-dependent | The default. If you are not running more than a few million tokens a month, the hardware never pays back. |
| 4× M5 Ultra Mac Studio | Four Mac Studio M5 Ultra, 512 GB each | ≈493 GB NVFP4, split across four machines | 36+ tokens/s (Apple, 8,192 context) | The best desktop story, and still a cluster. Two machines work but are slower; four are what Apple measured. |
| 4-GPU Blackwell node | One node, four NVIDIA Blackwell GPUs | ≈537 GB NVFP4 checkpoint | Not published by Meta | The vendor's intended self-host path. Less fuss than four Macs, more money, no desktop. |
| 8× B300, BF16 | Eight GPUs | ≈1.65 TB of weights, budget ≈2 TB VRAM | Not published by Meta | The reference-precision route. Only makes sense if you are fine-tuning or need the full format. |
| 2× DGX Spark | Two GB10 nodes, ≈121 GB each | 194.6 GiB IQ1_M across the pair | 6.76–7.01 tokens/s decode, 263 tokens/s prefill | The cheapest NVIDIA route we have a measurement for. The quant is aggressive and the context is 32K. |
| 128 GB MacBook, SSD streaming | One M5 Max MacBook, 128 GB | 99.2 GB peak; 1.9 GB of SSD traffic per token | ≈4.4 tokens/s decode | A demonstration, not a daily driver. It runs; the SSD is busy about 70% of the time. |
meta-models/meta-oss-cookbook; machinelearning.apple.com; two repository READMEs Verified 2026-10-17 Hosted price from OpenRouterNo speed figure is ours The Mac route, in detail
Apple's post is the most useful document of the launch week because it publishes the method
alongside the number. It evaluates Muse Spark Open in NVFP4 at 8,192 and 65,536 tokens of
context, using mlx_lm.benchmark, and reports the result for three parallelisation
schemes: pipeline over two Macs, tensor parallel over two, and tensor parallel over four.
The headline is more than 36 tokens per second on four M5 Ultra machines. The reason it works is bandwidth: 1.2 TB/s of unified memory per M5 Ultra, against 819 GB/s on the previous generation. Decode is memory-bandwidth bound, so that difference is most of the speed difference. Power stays around ≈300 W on M5 Ultra, ≈200 W on M3 Ultra, which is the argument for putting one under a desk rather than in a rack.
Two machines are enough to hold the model, and Meta's post describes the two-machine pipeline as "extremely simple to implement". But the measured number everyone quotes is the four-machine tensor-parallel configuration, and a reader budgeting a purchase should not assume the two-machine configuration reaches the same throughput. The weights fit; the speed is a separate question.
The NVIDIA routes
Meta's cookbook lists two checkpoint formats and the hardware each implies: BF16 at about 1.65 TB of weights, served on eight GPUs such as 8× B300 with roughly 2 TB of VRAM budgeted for cache and overhead; and NVFP4 at about 537 GB, served on a single four-GPU Blackwell node. Those are the vendor's routes, and neither comes with a published tokens-per-second figure.
The community filled that gap within a day. muse-spark-open-2x-dgx-spark documents two DGX Spark nodes holding a 194.6 GiB IQ1_M quant over llama.cpp RPC, with a measured 6.76–7.01 tokens per second at decode and 263 tokens per second at prefill. It is the cheapest NVIDIA route with a number attached, and the tradeoff is written into the quant: 32K context and an aggressive 1-bit format.
The laptop route, and why it is a demo
mlx-moe-ssd-streaming is the most interesting experiment of the first 48 hours. It keeps 16.9 GB of dense weights resident, holds a 25% cache of experts in 80 GB, and streams the rest from the SSD. On an M5 Max with 128 GB it reaches about 4.4 tokens per second, with 1.9 GB of SSD traffic per token and the disk busy about 70% of the time.
That is genuinely running the model, and it is not a workstation. The author is explicit about the bottleneck: an 84% expert-cache hit rate still produces a disk-bound token because the misses are what cost time, and more cache is the thing that helps most. Read the repository before buying a laptop for this; the result is a proof of concept on hardware that was already there.
Does any of this pay back?
At $1.25 input and $4.25 output per million tokens, a hosted endpoint is cheap until it is not. The crossover depends on volume and on what you count: four Mac Studios are four workstations, and a Blackwell node is a data-centre purchase, so the hardware cost is only comparable to the API bill if you amortise it over a lot of tokens and add power, cooling and somebody's time.
The honest reasons to self-host are not price per token. They are data residency, a workload that needs a fine-tuned checkpoint, latency you control, or a licence review that forbids sending prompts to a provider. If none of those apply, the hosted route is the one to take, and the local routes are worth reading about rather than buying.
What this page cannot tell you
We have run none of these routes. The Mac numbers are Apple's, the NVIDIA checkpoint sizes are Meta's, and the DGX Spark and SSD-streaming numbers are their authors'. Throughput on your workload, the effect of a longer context on memory, and the real electricity bill are all open questions. Treat every "tokens per second" above as a measurement of one setup, not a promise about yours.
Meta's cookbook → Apple's MLX post The hosted route Every build we hold