Model comparison

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 vs DeepSeek-V4-Flash-0731

These two publish different quantization sets; the table below has the exact sizes.

From the file· summed bytes, KV per layer

Side by side

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16DeepSeek-V4-Flash-0731
Parameters561B304B
Architecturenemotron_h_moedeepseek4
Layers43
Native context262,1441,048,576
Mixture of expertsyes, 512 expertsyes, 256 experts
Quantizations published1511
Smallest quantization175.12 GiB76.87 GiB
Q4_K_M
Licenceothermit

KV cache by context

the term that decides long-context viability
ContextNVIDIA-Nemotron-3-Ultra-550B-A55B-BF16DeepSeek-V4-Flash-0731Ratio
4,0960.06 GiB
8,1920.06 GiB
16,3840.06 GiB
32,7680.06 GiB
65,5360.06 GiB
131,0720.06 GiB