DeepSeek V4 Flash

DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts model from DeepSeek with a 1M-token context window, built for fast inference, high-throughput workloads, reasoning, coding, and agent workflows.

Reasoning
Tool Calling
Structured Output

Added Apr 24, 2026

Model weights

Pricing

Auto routing · per 1M tokens

Input: $0.14
Output: $0.28
Cache read: $0.0028

Compare provider prices

Specifications

Context window: 1M
Max output: 384K
Parameters: 284B / 13B; Total / active
Avg output (7d): 166 tokens; Longer than 13% of models

Benchmarks

Artificial Analysis

LMArena

Design Arena

Sourced from Artificial Analysis.

Intelligence Index

34.3

Better than 88% of models compared

Coding Index

69.1

Better than 80% of models compared

Agentic work

AutomationBench-AA

Workflow automation with guardrail penalties

54.0%

Better than 72% of models compared

AA-Briefcase

Agentic knowledge work (Elo)

1259 Elo

Better than 66% of models compared

GDPval-AA v2

Economically valuable tasks (Elo)

1438 Elo

Better than 79% of models compared

Document reasoning

GDP.pdf

Professional PDF reasoning: all-pass rate

11.0%

Better than 43% of models compared

AA-LCR v1.1

Long context reasoning with updated grading

79.7%

Better than 85% of models compared

MLCR-AA

Medical long-context reasoning

13.3%

Better than 43% of models compared

Reasoning

HLE

Humanity's Last Exam

38.6%

Better than 86% of models compared

IFBench

Instruction-following benchmark

47.2%

Better than 56% of models compared

CritPt

Research-level physics reasoning

0.3%

Coding

Terminal-Bench v4.0

Practical coding and terminal tasks

12.1%

Better than 64% of models compared

SciCode

Python programming for scientific computing

50.3%

Better than 54% of models compared

Knowledge

AA-Omniscience Accuracy

Proportion of correctly answered questions

26.0%

AA-Omniscience Hallucination Rate

Rate of incorrect answers among non-correct responses

94.5%

Legacy benchmarks

GPQA Diamond (legacy)

Graduate-level scientific reasoning

90.8%

Better than 91% of models compared

Terminal-Bench Hard (legacy)

Agentic coding and terminal use

34.1%

Better than 79% of models compared

T²-Bench Telecom (legacy)

Conversational AI agents in dual-control scenarios

94.4%

Better than 92% of models compared

AA-LCR (unversioned / legacy)

Long context reasoning evaluation

79.7%

Better than 84% of models compared

Last updated Oct 14, 2026