Stable production successor to V3.2-Exp, combining DeepSeek Sparse Attention with a scaled reinforcement-learning framework and agentic tool-use training.

Benchmarks

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

benchmark	published	measured
SWE-bench	73.1	70
HLE	25.1	—

Lineage

↑ finetune DeepSeek V3.2-Exp

view in full graph →

Metadata

org	DeepSeek
released	2025-12-01
params	685B
license	mit
hf id	deepseek-ai/DeepSeek-V3.2