Atlas

Models

← All models

DeepSeek-V4-Pro

DeepSeek · 2026-04-24 · 1.6T parameters

DeepSeek-V4-Pro is the larger open-weight Mixture-of-Experts model in DeepSeek's V4 preview release, with 1.6T total and 49B activated parameters. Released April 24, 2026, it supports a 1M-token context and Non-Think, Think High, and Think Max reasoning modes, and uses hybrid Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Its weights are released under the MIT license.

Benchmark scores

BenchmarkScore
A-Fantasia56.7
A-Fantasia Backwards Spelling45.0
A-Fantasia Chess44.0
A-Fantasia Cube Rotation81.0
AA-Briefcase Elo931
AA-LCR66.3
AA-Omniscience Index-9.7
Agent Red Teaming k=100 (CAIS)90.9
Agents' Last Exam (ALE)12.4
Agents' Last Exam ALE-CLI Pass Rate14.1
Agents' Last Exam ALE-CLI Score30.8
Agents' Last Exam Full-Spectrum Pass Rate10.9
Agents' Last Exam Full-Spectrum Score23.8
Agents' Last Exam Last-Exam Pass Rate0.0
Agents' Last Exam Last-Exam Score2.5
Agents' Last Exam Near-Term Pass Rate19.9
Agents' Last Exam Near-Term Score43.8
Agents' Last Exam Overall Score27.6
Agents' Last Exam Unlicensed Full-Spectrum Pass Rate11.0
Agents' Last Exam Unlicensed Full-Spectrum Score24.9
Agents' Last Exam Unlicensed Last-Exam Pass Rate0.0
Agents' Last Exam Unlicensed Last-Exam Score3.2
Agents' Last Exam Unlicensed Near-Term Pass Rate19.9
Agents' Last Exam Unlicensed Near-Term Score43.8
Agents' Last Exam Unlicensed Overall Pass Rate13.4
Agents' Last Exam Unlicensed Overall Score29.7
AHK-Eval75.0
AHK-Eval Algorithms66.7
AHK-Eval Data Structures33.3
AHK-Eval Date and Time50.0
Loading Atlas data…