0

DeepSeek V4 Pro 0423

Open weights

DeepSeek's April 2026 next-gen open-weights flagship - 1.6T-total / 49B-active MoE with 1M context and DeepSeek Sparse Attention.

Open
Publisher
DeepSeek
Family
Deepseek
Params
1.6T total / 49B active
Context
1.0M
Input
Output
Providers
deepseektogetherfireworksself-host
License
mit
Released
24 Apr 2026
Canonical
api-docs.deepseek.com/news/news260424

Cite

Notes

Only stored in your browser.

Attribution

Benchmark scores
AALiveBenchVaultmatharenaVals AIEQ-BenchLech Mazur
Attribution policy →
Context
1.0M
Input
$0.43
/1M tok
Output
$0.87
/1M tok
Blended
$0.54
#307/ 461/1M · 3:1
Speed
64
#81/ 461tok/s
Cache read
$0.09
0.20x input

Price & speed via Artificial Analysis · ranked across tracked models · caching via OpenRouter

Reasoning effortIndexSpeed
Reasoning, max45.364 t/s
Reasoning, high43.766 t/s
Non-reasoning31.974 t/s

Intelligence Index per reasoning-effort setting, via Artificial Analysis · pricing identical across tiers

Reported on 27 evals across 12 domains - best 91.2% on τ²-bench (Tau²-bench) · #31 of 321

Reported eval scores

27
SWE-bench80.6
τ²-bench (Tau²-bench)91.2
LiveBench - Language78.1
Terminal-Bench (Hard)36.4
LiveBench - Math90.7
LiveBench - Data Analysis74.5
LiveBench - Reasoning82.7
LiveBench - Coding70
IFBench45.8
LiveBench - Instruction Following62.4
LiveBench76.4
Humanity's Last Exam (HLE)8.2
Vals Index42.9
SciCode42.4
GPQA Diamond71.7
EQ-Bench Longform Creative Writing75.6
EQ-Bench Creative Writing v31552.1
Creative Story-Writing Benchmark47
MathArena76.5
Vibe Code Bench v1.149.9