DeepSeek-V4-Pro

Claims and evidence

• Released 24 April 2026 ● Primary source[1] [4]Date: 2026-04-24 [3]General Information: Release date

• Status AvailablePreview weights remain downloadable from the deepseek-ai Hugging Face organisation. In the DeepSeek API the model name deepseek-v4-pro has served the GA release DeepSeek-V4-Pro-0813 since 2026-08-13. ● Primary source[2]Model Downloads; License [4]Date: 2026-08-13
• Replaced by DeepSeek-V4-Pro-0813In the DeepSeek API, app and web: the GA release took over the model name deepseek-v4-pro on 2026-08-13; its card says it supersedes the preview. ● Primary source[4]Date: 2026-08-13 [5]Introduction

• Successor of DeepSeek-V3.2 EditorialEditorial link along the DeepSeek-V line: the V4 card and report compare efficiency and base-model results with DeepSeek-V3.2, and the report abstract speaks of unnamed predecessors, but no source names V4-Pro the successor of V3.2. ● Primary source[2]Introduction; Base Model evaluation [6]Abstract

• Change · Architecture Introduces hybrid attention (Compressed Sparse Attention plus Heavily Compressed Attention), Manifold-Constrained Hyper-Connections and training with the Muon optimizer.Compared with DeepSeek-V3.2 ● Primary source[2]Introduction
• Change · Efficiency At a 1M-token context it needs about 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2, according to DeepSeek.Compared with DeepSeek-V3.2 ● Primary source[2]Introduction
• Change · Size 1.6T total and 49B activated parameters, against 671B total and 37B activated for DeepSeek-V3.2-Base in the card's base-model table.Compared with DeepSeek-V3.2 ● Primary source[2]Base Model evaluation
• Change · Context length Supports a 1M-token context, which DeepSeek made the default across its official services with V4.Compared with DeepSeek-V3.2 ● Primary source[1]

• Input text ● Primary source[3]Model properties: Modalities
• Output text ● Primary source[3]Model properties: Modalities
• Feature Reasoning modeNon-think, Think High and Think Max modes; the API offers thinking and non-thinking modes. ● Primary source[2]Instruct Model: reasoning effort modes [1]API is Available Today [3]Model properties: Modalities
• Feature Long context ● Primary source[1] [2]Introduction
• Open weights YesWeights and code under the MIT License. ● Primary source[2]License [3]Methods of distribution and licenses [1]
• Context window 1M tokens ● Primary source[2]Model Downloads [1] [3]Model properties: Context length
• Parameters 1.6T total / 49B activeMixture-of-Experts: 1.6T total parameters, 49B activated per token. ● Primary source[1] [2]Model Downloads [3]Model properties: Total model size
• Access API, consumer app, Open-weights download ● Primary source[1] [3]Methods of distribution and licenses

Lineage

Variants

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from DeepSeek on AI Radar →