Meta’s Muse Spark Open really can run at home. Hugging Face reports 18 output tokens per second from a 2-bit build on a 512GB M3 Ultra Mac Studio, and llama.cpp has merged support for the model’s text, image and audio inputs.[2][7]
That is a remarkable local-AI result.[2] It is also a rather elastic use of the word “desktop.”
Meta launched the model on October 15 with 816 billion total parameters, 42 billion active for each token and a context window of up to one million tokens. The company said it brings “leading agentic capabilities” to a desktop for the first time, using two M5 Ultra Mac Studios with 512GB of unified memory apiece.[1]
The narrow verdict is that Meta’s run-it-at-home claim holds up. The broader suggestion that a frontier-class agent has become accessible to ordinary home users does not, at least not yet. The weights are enormous, the tested local build sacrifices precision, the advertised maximum context adds another mountain of memory, and the exact Apple configuration was still marked “coming late October” when checked on October 26.[2][5]
Active parameters do not determine what fits
Muse Spark Open is a mixture-of-experts model. A router chooses eight of 256 experts for each token, which cuts the arithmetic to roughly 42 billion active parameters. All 256 experts still have to live somewhere. Hugging Face puts the BF16 checkpoint at about 1.5TB and says deployment needs the memory capacity of an 816-billion-parameter model even though generation behaves more like a 42-billion-parameter workload.[2]
That distinction is where the friendly desktop headline meets the hardware bill.[1][2]
Meta’s model card calls for at least 700GB of GPU memory for its NVFP4 checkpoint or 2TB for BF16. It separately points to two M5 Ultra Mac Studios for MLX.[3] The public ggml-org repository offers a more aggressive route for llama.cpp: its 22 Q2_K shards total 304.08GB, while Q4_K_S totals 470.07GB. Those figures come from adding the exact file sizes in the repository API; multimodal projector files add roughly another 2.18GB to 4.10GB.[4]
A 512GB Mac can plausibly hold the Q2 model with useful context and runtime overhead.[2][4] Hugging Face says its imatrix-calibrated Q2 build used 304GB and produced 18 tokens per second on an M3 Ultra with 512GB.[2] That is conversational speed for a very large model.[2] It is one partner’s measurement, however, not an independent HighVRAM test, and a 2-bit quant is not equivalent to Meta’s higher-precision checkpoint.[2][3]
Q4 is a tighter proposition.[4] A 470GB model leaves only about 42GB of nominal unified memory before macOS, llama.cpp, the multimodal projector and the context cache take their shares.[4] Buying 512GB does not mean an application gets every last byte. This is less a comfortable setup than a memory Tetris tournament with a five-figure entry fee.
The million-token context needs its own machine-sized allowance
The one-million-token context is real in the model configuration, but it should not be folded into the desktop claim for free.
Hugging Face documents 126,976 key-value cache values per token. At BF16, its full 1,048,576-position window consumes about 266GB.[2] Using one million tokens for a round number gives 253.95GB. Add that to the 304.08GB Q2 model and the total reaches roughly 558GB before operating-system or runtime overhead.
One 512GB Mac therefore cannot hold the tested Q2 build and a full million-token BF16 cache at the same time.[2][4][5] A shorter 128,000-token context needs about 32.5GB by the same calculation, which is far more plausible.[2] That is also the context limit in Hugging Face’s sample vLLM configuration.[2]
Two 512GB Macs solve the capacity problem on paper. Apple documents up to 512GB of unified memory and 1.2TB/s of memory bandwidth per M5 Ultra Mac Studio.[5] Meta says MLX can make two of them hold the complete model.[1] What the launch material does not establish is that every advertised capability, at maximum context, runs at a speed and quality worth buying two top-end systems to obtain.[1][2]
Available weights, unavailable shopping cart
Meta’s “available today” language is accurate for hosted APIs and weight access.[1] The official Hugging Face page is gated and asks for a legal name, birth date and organization details, while a public ggml-org repository exposes the quantized llama.cpp artifacts.[3][4] The model is open-weight, not open source.[3] The Muse license requires attribution, incorporates a usage policy and sets a separate-agreement condition for certain model-as-a-service businesses above $20 million in revenue.[3]
The hardware story was less immediate. Apple’s store listed the M5 Ultra Mac Studio from $5,499 and still described the 512GB option as “coming late October” on October 26.[5][6] Two base M5 Ultra machines start at $10,998, but base memory is not the configuration Meta specifies.[1][6] The retrieved store page did not expose a final price for two 512GB systems, so any precise total would be guesswork.[6]
This leaves the availability claim split in two. The model exists, the weights exist, and a merged llama.cpp implementation exists.[3][4][7] The exact new-hardware path promoted in Meta’s announcement was not yet a finished purchase when checked.
What remains a claim
Meta calls Muse Spark Open the most capable US open-weight model to date and presents it as a strong long-horizon coding agent.[1] The company publishes benchmark results and demonstrations, but the sources inspected for this audit did not provide independent testing of its multi-file coding reliability, tool use or one-million-token behavior.[1][3] Hugging Face’s 18-token-per-second result supports local feasibility, not Meta’s capability ranking.[2]
That does not make the release empty marketing. Shipping an 816B MoE model in formats that llama.cpp can run on one 512GB Mac is substantive.[2][4] It moves an extreme model from data-center-only to workstation-possible, gives owners control over prompts and weights, and keeps multimodal input on premises.[4][7]
For the small group that already owns a 512GB Apple-silicon machine, the Q2 build is worth trying now.[2][4] Everyone else should wait for independent agentic tests and final 512GB M5 Ultra pricing before treating Meta’s desktop as a buying recommendation.[1][5][6] A hosted endpoint or a smaller local model will be the saner choice for most homes.[1][2]
Muse Spark Open crossed the boundary into home hardware.[2][4] It did not make the boundary cheap.[5][6]