Is DeepSeek-V3.2 or DeepSeek-V4-Flash-0731 smaller?

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

Which uses less memory at long context?

DeepSeek-V4-Flash-0731 — its KV cache at 32K context is 34.0x smaller, because of how its attention layers cache.

Model comparison

DeepSeek-V3.2 vs DeepSeek-V4-Flash-0731

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: DeepSeek-V4-Flash-0731's KV cache at 32K is 34.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	DeepSeek-V3.2	DeepSeek-V4-Flash-0731
Parameters	685B	304B
Architecture	deepseek2	deepseek4
Layers	61	43
Native context	163,840	1,048,576
Mixture of experts	yes, 256 experts	yes, 256 experts
Quantizations published	29	11
Smallest quantization	138.82 GiB	76.87 GiB
Q4_K_M	376.71 GiB	—
Licence	mit	mit

KV cache by context

the term that decides long-context viability

Context	DeepSeek-V3.2	DeepSeek-V4-Flash-0731	Ratio
4,096	0.27 GiB	0.06 GiB	4.26×
8,192	0.54 GiB	0.06 GiB	8.51×
16,384	1.07 GiB	0.06 GiB	17.02×
32,768	2.14 GiB	0.06 GiB	34.05×
65,536	4.29 GiB	0.06 GiB	68.09×
131,072	8.58 GiB	0.06 GiB	136.19×