DeepSeek · DeepSeek-V4

DeepSeek-V4-ProFlagship

CurrentVerified · Jun 27, 2026

DeepSeek's flagship V4 MoE model (1.6T total / 49B active params), released April 2026 — frontier reasoning, coding and agentic performance with a 1M-token context. Open weights.

profile normalized against the 70-model field

Context window· 1M of 10M10%

Max output· 384K of 384K100%

Output speed—

Affordability· $0.87 / Mtok out100%

Capability breadth· 2 of 1118%

Capability switches · 2 of 11

Reasoning mode

Tool / function use

Streaming

JSON mode

Structured outputs

Prompt caching

Fine-tuning

Web search

Code execution

Vision input

Audio input

Specifications

Every value carries a primary source and a verification date.

Capacity

Context window

Max output

384K

Pricing

Input $/Mtok

$0.435 USD / 1M tokens

Cached input $/Mtok

$0.003625 USD / 1M tokens

Output $/Mtok

$0.87 USD / 1M tokens

Capabilities

Reasoning mode

Yes

Tool / function use

Yes

Vision input

API

API model ID

deepseek-v4-pro

General

Release date

April 24, 2026

Benchmarks

Sourced evaluation scores, each verified against its primary source.

MMLU-Pro

MMLU-Pro (EM) | 89.1 | 87.5

87.5 %Verified

GPQA Diamond

GPQA Diamond (Pass@1)

90.1 %Verified

SimpleQA-Verified

SimpleQA-Verified (Pass@1)

57.9 %Verified

Humanity's Last Exam

HLE (Pass@1)

37.7 %Verified

LiveCodeBench

LiveCodeBench (Pass@1)

93.5 %Verified

Codeforces

Codeforces (Rating)

3206 ratingVerified

HMMT 2026 Feb

HMMT 2026 Feb (Pass@1)

95.2 %Verified

IMOAnswerBench

IMOAnswerBench (Pass@1)

89.8 %Verified

Apex Shortlist

Apex Shortlist (Pass@1)

90.2 %Verified

Terminal Bench 2.0

Terminal Bench 2.0 (Acc)

67.9 %Verified

SWE-bench Verified

SWE Verified (Resolved)

80.6 %Verified

SWE-bench Pro

SWE Pro (Resolved)

55.4 %Verified

SWE-bench Multilingual

SWE Multilingual (Resolved)

76.2 %Verified

BrowseComp

BrowseComp (Pass@1)

83.4 %Verified

Chinese-SimpleQA

Chinese-SimpleQA (Pass@1)

84.4 %Verified

Loading…