Muse Spark Open Runs on a 512GB M3 Ultra—But Not the Official NVFP4
Published Muse Spark Open tests used 304GB Q2_K and 344GB oQ3e builds on a 512GB M3 Ultra. They are not Meta’s official 700GB NVFP4 path.
Evidence-first local AI model sizing, hardware comparisons and deployment analysis. Capacity and published specifications are kept separate from measured performance.
Published Muse Spark Open tests used 304GB Q2_K and 344GB oQ3e builds on a 512GB M3 Ultra. They are not Meta’s official 700GB NVFP4 path.
Muse Spark Open’s official four-device SGLang reference uses GB300-class hardware, TP4 and EP4. See its 1,152GB arithmetic and evidence limits.
Muse Spark Open’s prevalidated 1,048,576-token endpoint uses eight H200 GPUs totaling 1,128GB. Here is what that evidence does—and does not—prove.
The published 8×RTX PRO 6000 Muse Spark Open recipe totals 768GB and targets 131,072-token context. Here is what it validates—and what it does not.
Four B200 GPUs total 720GB, only 20GB above Muse Spark Open’s stated minimum. The arithmetic fits, but the official reference is a full B200 node.
Muse Spark Open has prevalidated eight-RTX-PRO and eight-H200 hosted paths. Compare their 768GB and 1,128GB pools, context targets and evidence limits.
Meta’s Muse Spark Open NVFP4 files total 533.011GB, yet official guidance says at least 700GB VRAM. Here is why 640GB is still below the line.
Eight H100 80GB GPUs total 640GB, leaving 44.822GB after Kimi K2.5’s official files—just 5.603GB per GPU under perfect balance.
Four B200 GPUs provide 720GB and 124.822GB of file margin for Kimi K2.5; three MI325X accelerators provide 768GB and 172.822GB. Capacity is only the first gate.
Kimi K2.5’s official checkpoint is 595.178GB. Here is why 600GB is only a paper fit, 640GB remains tight, and 768GB is a safer planning tier.