DeepSeek-VL2 vs DeepSeek V4-Flash vs DeepSeek V4-Pro

Best cell per row highlighted. Null means undisclosed — never counted as zero.

DeepSeek
DeepSeek-VL2GA
7.2
AI Panel
DeepSeek
DeepSeek V4-Flashpreview
8.7
AI Panel
DeepSeek
DeepSeek V4-Propreview
8.5
AI Panel

Identity & lifecycle

ProviderDeepSeekDeepSeekDeepSeek
Family / tierDeepSeek-VL · VLDeepSeek V4 · FlashDeepSeek V4 · Pro
StatusGApreviewpreview
Released2024-12-122026-04-232026-04-23
Knowledge cutoff2024-102026-022026-02

Architecture & context

Context window4K1M1M
Max output tokens4K384K384K
Input modalitiestext, imagetexttext
Reasoning modenoneoptionaloptional
Open weightsYesYesYes

Pricing (per Mtok)

Input$0.14$0.435
Output$0.28$0.87
Cached input$0.0028$0.0036
Batch input
Free tierYesYesYes

Speed

Output speed (tok/s)
Time to first token (s)
Latency tierfastfastmedium

Trust & deployment

Trains on inputsNoYesYes
Licensecustom-deepseek-model-licenseMITMIT
Clouds
Compliance

AI Panel scoring

Unified score7.28.78.5
Decision Maker7.58.58
Domain Strategist78.58.5
Finance Lead8.59.79.5
Domain Practitioner7.58.58.5
Power User6.587.5
Skeptic77.57.5
Value score8.89.99.7

Benchmarks

MMLU-Pro86.287.5
GPQA Diamond88.190.1
SWE-bench Verified7980.6
HumanEval76.8
LiveCodeBench91.693.5
MMMU51.1
MRCR Long Context78.783.5
SimpleQA34.157.9
Humanity's Last Exam34.837.7
LMArena Coding Elo1287
Artificial Analysis Index52

Scores link to their sources. Missing cells mean the vendor hasn't disclosed a result — honesty over padding.