Skip to content
EN ES
DeepSeek

DeepSeek V4 Pro

Released 2026-04 · hybrid · 1.0M tokens · 8 benchmarks · Open weight
HuggingFaceGitHub

Editorial notes

Preview lanzado 24 abril 2026. 1.6T total / 49B activos. Hybrid Attention (CSA + HCA). 3 niveles de effort: Non-think / Think High / Think Max. 27% FLOPs y 10% KV-cache vs V3.2 en context 1M. Codeforces 3206. MIT license. Pricing $1.74/$3.48 per MTok (cache $0.015, -99%). Reemplaza deepseek-chat y deepseek-reasoner. Scores de la columna Max del model card oficial deepseek-ai/DeepSeek-V4-Pro (GPQA-D 90.1, HLE 37.7 sin tools / 48.2 con tools, LiveCodeBench 93.5, SWE-V 80.6, SWE-Pro 55.4, MMLU-Pro 87.5). AIME-2025 87.5 REMOVIDO en auditoria 2026-06-08: el card NO reporta AIME — ese 87.5 era MMLU-Pro mal etiquetado (ya cargado correcto).

Spec sheet

Company
DeepSeek
Country
CN
Type
hybrid
Release
2026-04
Context
1.0M tokens
Params total
1.6T
Params active (MoE)
49B
License
mit
Quants
BF16, Q8_0, Q5_K_M, Q4_K_M, Q3_K_M
Pricing (deepseek)
$0.435/$0.87/M
Slug
deepseek-v4-pro

Quick install

2 tools
ollama.com 
ollama run deepseek-v4:pro

Note: Q4_K_M requiere multi-GPU/multi-node

Benchmarks (8)

Reasoning 3

Coding 4

Math 1

Cite this model
BibTeX · APA

BibTeX 

@misc{frontier-deepseek-v4-pro,
  title  = {DeepSeek V4 Pro},
  author = {{DeepSeek}},
  year   = {2026},
  note   = {Frontier Benchmarks AI atlas. Accessed 2026-06-17},
  url    = {https://frontierbenchmarks.com/models/deepseek-v4-pro}
}

APA

DeepSeek (2026). DeepSeek V4 Pro [Large language model]. Frontier Benchmarks AI. Retrieved 2026-06-17, from https://frontierbenchmarks.com/models/deepseek-v4-pro

Citation reflects the atlas page, not the original model paper. For the paper, see the "Resources" section above.

Battle vs another model ← All models More from DeepSeek Methodology