DeepSeek
DeepSeek

DeepSeek V4 Flash

Release Date 2026-04-24(Knowledge cutoff 2025-05)1.0M IN · 384K OUTOpen Model(MIT)API available

DeepSeek V4 Flash is an open-weight language model for long text, reasoning and tool use, using a smaller configuration than its same-generation Pro counterpart. DeepSeek described attention changes that reduce long-input computation and memory demands over the earlier generation, with thinking and non-thinking modes for repeated workloads. Self-managed deployment is possible, but sparse expert activation does not eliminate the hardware and memory needed to store the full model. Current AIB comparisons put automation in the upper tier, documents in the middle of the pack and coding on the weaker side. The original model is retired on the official API, and its legacy Flash name routes to a successor, so distinguish original-weight evaluations from current service results and verify the model served, cost and output.

API cost

Market price (median)
$0.135IN$0.28OUT
Price range19 providers
$0.044~$0.44IN$0.132~$1.32OUT
IN
OUT
10/2311/22
SourceOpenRouterLast updated

Real-world usage

No.217 places0.8% share · 2.2T tokens/wk
12T8/31
12T9/7
11T9/14
7.8T9/21
6.0T9/28
4.9T10/5
4.7T10/12
4.0T10/19
3.9T10/26
3.3T11/2
3.3T11/9
2.2T11/16
Top use cases
  • Remembering & organizing conversations17.7%
  • Roleplay & fiction12.7%
  • Content writing10.2%
SourceOpenRouterLast updated

Output Speed

Very fastper 2,000 tokens
113tok/s
First Output 0.5stotal: 18.2s
SourceOpenRouterLast updated

User Comparisons

Writing Style
Answer length
Shorter side
bottom 35%
Vocabulary diversity
Plainer side
bottom 35%
Sentence rhythm
Steadier side
bottom 39%
Language mixing
High
top 30%
List usage
Occasionally
11% of lines
Heading usage
Occasionally
7% of lines
Reasoning share
Moderate
15% of total
Response Speed
instability ±347%
First response 9.2sTotal 26.3s
50 tok/s
12.4ms per token
SourceAIB Inc.Last updated

Expert Performance Evaluation

143665.534.3

More about Arena Text

39.4%95.0%69.0%

More about τ³-Banking

79.7%143849.5%61.8%28.1%30.3%41.4%68.08.3%

More about Long Context Reasoning

143933.6%73.2%45.1%4.3%

More about Arena Text Factuality

12.1%148437.6130652.6%78.7%35.6%69.150.3%18.8%63.0%147169.20.0%74.7%

More about Terminal-Bench 4.0

1416145379.2%1407142870.180.4%140563.1

More about Arena Writing, Literature & Language

38.6%16.6%142490.8%94.4%79.757.5%24.4%56.0%88.4%1457143470.689.0%61.4%61.1%32.0%34.0%90.9%41.7

More about Humanity's Last Exam

Specialized fields

13941397145914571459144895.8%93.9%94.9%85.1%27.1%89.4%58.9%81.7%86.0%80.0%70.7%

More about Arena Korean

OddBench

39.0%Reasoning effort: Extra High89.6%414

By effortNone 20.0%Extra High 39.0%

More about BullshitBench v2

See each evaluation’s detail page for its source and last updateData corrections & removal requests