Is DeepSeek-V4-Flash or DeepSeek-V2.5 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 33.5x smaller, because of how its attention layers cache.

Model comparison

DeepSeek-V4-Flash vs DeepSeek-V2.5

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 33.5× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	DeepSeek-V4-Flash	DeepSeek-V2.5
Parameters	291B	236B
Architecture	deepseek4	deepseek2
Layers	43	60
Native context	1,048,576	163,840
Mixture of experts	yes, 256 experts	yes, 160 experts
Quantizations published	12	42
Smallest quantization	76.87 GiB	44.14 GiB
Q4_K_M	—	132.67 GiB
Licence	mit	other

KV cache by context

the term that decides long-context viability

Context	DeepSeek-V4-Flash	DeepSeek-V2.5	Ratio
4,096	0.06 GiB	0.26 GiB	4.19×
8,192	0.06 GiB	0.53 GiB	8.37×
16,384	0.06 GiB	1.05 GiB	16.74×
32,768	0.06 GiB	2.11 GiB	33.49×
65,536	0.06 GiB	4.22 GiB	66.98×
131,072	0.06 GiB	8.44 GiB	133.95×