Atlas

Models

← All models

DeepSeek-V2

DeepSeek · 2024-05-06 · 235.7B parameters

DeepSeek-V2 is DeepSeek's pretrained base checkpoint released May 6, 2024; its downloadable weights contain exactly 235,741,434,880 parameters, while official architecture figures are rounded to 236B total and 21B activated per token. It was pretrained on a reported 8.1T-token corpus with MLA and DeepSeekMoE (2 shared and 160 routed experts per MoE layer, 6 routed experts selected per token), then extended with YaRN from 4K to documented 128K context support; its released configuration declares 163,840 maximum positions.

Benchmark scores

BenchmarkScore
ARC AI292.2
BBH (BIG-Bench Hard)78.8
Capability119.3
Epoch Capabilities Index (ECI)124
HellaSwag (Unspecified Scoring Protocol)87.1
MMLU78.4
PIQA83.9
TriviaQA80.0
WinoGrande86.3
Loading Atlas data…