Atlas

Models

← All models

DeepSeek-V4-Flash-Base

DeepSeek · 2026-04-24 · 292B parameters

DeepSeek-V4-Flash-Base is the pretrained base checkpoint behind DeepSeek-V4-Flash, published with the V4 preview release on April 24, 2026 in FP8 mixed precision. DeepSeek markets it at 284B total and 13B activated parameters; its released safetensors contain 292,021,347,282 stored values including the multi-token-prediction module and quantization scales. It shares the V4 hybrid CSA/HCA attention design and was pretrained on more than 32T tokens, with a 1M-token context.

Benchmark scores

BenchmarkScore
BBH (BIG-Bench Hard)86.9
BIG-Bench Extra Hard25.4
C-Eval92.1
Capability128.5
CLUEWSC82.2
CMATH93.6
CMMLU90.4
DROP88.6
GSM8K90.8
HellaSwag (Unspecified Scoring Protocol)85.7
HumanEval69.5
LongBench v244.7
MATH57.4
MGSM85.7
MMLU88.7
MMLU Pro68.3
MMLU-Redux89.4
MMMLU88.8
SimpleQA Verified30.1
TriviaQA82.8
WinoGrande79.5
Loading Atlas data…