howtospark
models / nemotron-3-puzzle-75b-a9b

Nemotron 3 Puzzle 75B-A9B

nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16
params
75.4B
active params
9.0B
sparsity
MoE
attention
Full attention
architecture
nemotron_h_puzzle
license
other
downloads
2.2k
likes
49
hf updated
Jul 7, 2026

Visualize

total size

each column is a layer — lit blocks are the experts that fire per token

bits / param
16-bit
footprint
140 GiB
sparks to fit
2×
block ≈
589M params
active paramsdormant expertsghost = precision traded awayblock fill ∝ bits kept · Nemotron 3 Puzzle 75B-A9B ≈ 75.353324544B params