Back to the index
DeepSeek · DeepSeek
No. 25

DeepSeek-V4 Flash

Cheap, fast open MoE — near-frontier value at a fraction of Pro's cost and size.

FastOpen weightsText, Image
Field notes
Context window
1Mtokens
Max output
384Ktokens
Input price
$0.14/ 1M tokens
Output price
$0.28/ 1M tokens
Cached input
$0.01/ 1M tokens
Modalities
Text, Image
Knowledge cutoff
Jan 2026
Released
Apr 2026
Parameters
284B
License
MIT
Approx. speed
Overview

What this model is

DeepSeek-V4 Flash is the lightweight sibling of V4 Pro (284B-parameter MoE, 13B active, MIT), built for high-volume, latency-sensitive work. It keeps the V4 generation's 1M-token context, native image input, and DeepSeek Sparse Attention while cutting price to roughly a third of Pro's and running noticeably faster. It is one of the cheapest capable multimodal long-context models available, ideal as a router target for the bulk of everyday traffic, stepping up to Pro or a frontier model only for the hardest jobs.

Strengths

  • Extremely cheap — among the lowest prices anywhere
  • Fast, low-latency decoding for high-volume work
  • 1M-token context and native image input
  • MIT open weights, self-hostable

Trade-offs

  • Well below the frontier on hard reasoning and coding
  • Not for long-horizon autonomous agentic work
  • Vision quality is basic
  • Smaller active-parameter budget caps peak quality

Best for

  • High-volume classification, extraction, and routing
  • Cheap long-context document processing
  • Low-latency chat and simple coding
  • Batch multimodal (image + text) workloads

Not ideal for

  • Complex autonomous coding
  • Deep multi-step frontier reasoning
Capabilities

Capability profile

Normalized 0–100 scores, comparable across the whole catalog.

Reasoning74Coding62Math82Writing74Knowledge76Speed82Agentic58Vision58Multilingual78Long Context84
Reasoning
74
Coding
62
Math
82
Writing
74
Knowledge
76
Speed
82
Agentic
58
Vision
58
Multilingual
78
Long Context
84
Benchmarks

How it scores

Public benchmark results, with independent third-party results where available. Bars normalize percentages to 100 and Elo ratings to a 1500 ceiling.

Independent index
Artificial Analysis Intelligence Index
Composite of ~9–10 independent evals · Artificial Analysis
40/ 100
BenchmarkResult
Artificial Analysis Intelligence Indexindependent
GeneralArtificial Analysis
40%
SWE-bench Verifiedindependent
CodingSWE-bench
66%
SWE-bench Proindependent
CodingSWE-bench Pro
38%
GPQA Diamondindependent
ReasoningIndependent aggregators
74%
AIME 2025
MathVendor-reported
85%
LiveCodeBenchindependent
CodingLiveCodeBench
66%
Alternatives

Comparable models

Models in a roughly similar class — worth weighing against this one.