DeepSeek V3.2-Exp

DeepSeek llm

An experimental update to V3.1-Terminus introducing DeepSeek Sparse Attention for more efficient long-context training and inference at roughly the same output quality.

overall ⌄

43.0

74 of 125 ranked

price · $/M tokens

self-hosted

open weights, no first-party API

to run it yourself ⌄

403 GB

minimum · 1.6 TB recommended

Benchmarks

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

benchmarkpublishedmeasured
MMLU-Pro85 —
GPQA79.9 —
AIME 202589.3 —
SWE-bench67.8 —
HLE19.8 —
BrowseComp40.1 —

Lineage

↑ finetune DeepSeek V3.1
↓ finetune DeepSeek V3.2
view in full graph →

Metadata

orgDeepSeek
released2025-09-29
params671B-A37B
licensemit
hf iddeepseek-ai/DeepSeek-V3.2-Exp