Is NVIDIA-Nemotron-3-Super-120B-A12B-BF16 or DeepSeek-V4-Flash smaller?

Both publish quantizations; compare the table for the exact sizes.

Which uses less memory at long context?

DeepSeek-V4-Flash — its KV cache at 32K context is 43.7x smaller, because of how its attention layers cache.

Model comparison

NVIDIA-Nemotron-3-Super-120B-A12B-BF16 vs DeepSeek-V4-Flash

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: DeepSeek-V4-Flash's KV cache at 32K is 43.7× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	NVIDIA-Nemotron-3-Super-120B-A12B-BF16	DeepSeek-V4-Flash
Parameters	124B	291B
Architecture	nemotron_h_moe	deepseek4
Layers	88	43
Native context	262,144	1,048,576
Mixture of experts	yes, 512 experts	yes, 256 experts
Quantizations published	45	12
Smallest quantization	43.19 GiB	76.87 GiB
Q4_K_M	80.14 GiB	—
Licence	other	mit

KV cache by context

the term that decides long-context viability

Context	NVIDIA-Nemotron-3-Super-120B-A12B-BF16	DeepSeek-V4-Flash	Ratio
4,096	0.34 GiB	0.06 GiB	5.46×
8,192	0.69 GiB	0.06 GiB	10.91×
16,384	1.38 GiB	0.06 GiB	21.83×
32,768	2.75 GiB	0.06 GiB	43.66×
65,536	5.50 GiB	0.06 GiB	87.32×
131,072	11.00 GiB	0.06 GiB	174.64×