Interactive planner for Muse Spark Open. Weights and config.json are from Hugging Face's public listing for Meta's two repositories; memory and bandwidth are as printed on NVIDIA and Apple pages. All read 18 October 2026. Derived means this site's arithmetic, and the ceilings are upper bounds, not benchmarks. Weights: 1,644 GB in BF16 and 533 GB in NVFP4, 3.08 times smaller (derived). Six machines hold 640, 768, 1,024, 1,128, 1,440, 13,400 GB. NVFP4 fits in 6 of them; BF16 fits only in NVIDIA GB200 NVL72 rack. Cache of one sequence in BF16 (derived): at 131,072 tokens 9.0 GB on the config.json basis of 16 full and 46 windowed layers (untested) and 33.3 GB on the post's formula of all 62 layers; at 1,048,576 tokens 69.1 and 266.3 GB. Bytes read per token (derived): BF16 84.0 GB and NVFP4 49.3 GB, of which 13.6 GB is 4 bit experts and 35.7 GB stays BF16. The naive reading of 42B at 4 bits, 21.0 GB, is not what the files deliver. With 8 x NVIDIA H200 and NVFP4 weights: 595 GB free after the weights; 8.6 sequences of 1,048,576 tokens fit on the config.json basis and 2.2 on the post's formula; the decode ceiling is 324 tokens per second, 122 on the post's formula (derived upper bounds, nothing run). Choose a machine, a weights format and a context length to change every figure.
Memory and bandwidth as the vendor pages print them (summed over the units where the page prints one unit), against the weights: NVFP4 533 GB, BF16 1,644 GB
Machine
Memory
Bandwidth
NVFP4 weights
BF16 weights
8 x NVIDIA H100 SXM
640 GB
26.8 TB/s
fits
does not fit
8 x NVIDIA RTX PRO 6000 Blackwell Server
768 GB
12.8 TB/s
fits
does not fit
2 x Mac Studio, M5 Ultra, 512 GB
1,024 GB
2.4 TB/s
fits
does not fit
8 x NVIDIA H200
1,128 GB
38.4 TB/s
fits
does not fit
NVIDIA DGX B200
1,440 GB
64 TB/s
fits
does not fit
NVIDIA GB200 NVL72 rack
13,400 GB
576 TB/s
fits
fits
Decode ceilings for the NVFP4 weights, in tokens per second (derived upper bounds: bandwidth divided by the bytes one token reads, one stream, nothing run)
Machine
Weights alone
With the cache at 131,072 tokens
With the cache at 1,048,576 tokens
At 1,048,576 tokens, the post's formula
8 x NVIDIA H100 SXM
544
460
226
one sequence does not fit
8 x NVIDIA RTX PRO 6000 Blackwell Server
259
219
108
one sequence does not fit
2 x Mac Studio, M5 Ultra, 512 GB
49
41
20
8
8 x NVIDIA H200
779
659
324
122
NVIDIA DGX B200
1,299
1,099
541
203
NVIDIA GB200 NVL72 rack
11,690
9,889
4,866
1,825
Cache of one sequence in BF16 (derived). The config.json basis assumes a serving stack frees entries outside the window of the 46 windowed layers, which was not tested.
Context in tokens
config.json: 16 full and 46 windowed layers
The post's formula: all 62 layers
2,048
0.52 GB
0.52 GB
8,192
0.92 GB
2.1 GB
32,768
2.5 GB
8.3 GB
131,072
9.0 GB
33.3 GB
262,144
17.6 GB
66.6 GB
524,288
34.7 GB
133.1 GB
1,048,576
69.1 GB
266.3 GB
Bytes stored and bytes read per decoded token (derived from the stated 42B active parameters)
Format
Stored (file listing)
Read per token (derived)
BF16
1,644 GB
about 84.0 GB, up to 89.4 GB on the upper reading
NVFP4
533 GB
about 49.3 GB, up to 54.6 GB on the upper reading (13.6 GB of FP4 experts and 35.7 GB of BF16)
Naive: 42B at 4 bits
not a real file
21.0 GB, which the files do not deliver
Nothing here is a benchmark. Real throughput on any of these machines is not stated. Batching, prefix caching, interconnect, speculative decoding and expert placement are not modelled.