DeepSeek V4-Pro

DeepSeek llm

DeepSeek's V4 flagship with a 1M-token context, combining Compressed Sparse Attention and Heavily Compressed Attention. Scores are from its Think Max extended-reasoning mode.

overall ⌄

65.6

20 of 125 ranked

price · $/M tokens

0.435 / 0.87

in / out, as of 2026-07-25

to run it yourself ⌄

960 GB

minimum · 3.8 TB recommended

Benchmarks

What its maker published, and what anyone else measured. Every figure links the document it came from.

benchmarkpublishedmeasured
MMLU-Pro87.5 —
GPQA90.1 —
SWE-bench80.6 —
SWE-bench Pro55.4 —
HLE37.7 —
BrowseComp83.4 —

Cheaper, and at least as good

Nothing is both cheaper and ranked as high. This model is on the value frontier — whatever you pay for it, you are paying for something.

Lineage

view in full graph →

Metadata

orgDeepSeek
released2026-04-22
params1600B-A49B
licensemit
hf iddeepseek-ai/DeepSeek-V4-Pro