DeepSeek V4 Pro

DeepSeek

DeepSeek V4 Pro (public preview): 1.6T-param MoE, 49B active/token, hybrid attention. 1M context, MIT licensed.

Text Generation Local Latest DeepSeek Family vV4 Pro

Documentation Back to Models

Parameters

1.6T

params

Context Window

1.0M

tokens

Max Output

384K

tokens

Input Price

per 1M tokens

Output Price

per 1M tokens

DeepSeek Family 15 models

The full DeepSeek line by generation — pricing and capabilities vary across the family.

V4 Pro Current Latest

Agentic reasoning, coding, long-context

1.0M

context

Apr 2026

V4 Flash

Cost-sensitive inference, coding, long-context

DeepSeek

General purpose, reasoning, code generation

Code generation, completion

Professional coding tasks

V3.2

Sparse attention, experimental

$0.27 / $1.1

in / out · 1M

Sep 2025

V3.1

Hybrid thinking, 40% better on SWE-bench

Reasoning, math, coding, analysis

Reasoning, math, coding

131K

context

Jan 2025

Reasoning, chain-of-thought

General purpose, MIT license

$0.27 / $1.1

in / out · 1M

Mar 2025

Coder

Code generation, debugging

$0.14 / $0.28

in / out · 1M

Capabilities

👁️

Vision

⚡

Function Calling

📋

JSON Mode

🌊

Streaming

💬

System Prompt

🖥️

Code Execution

🔍

Web Search

🔌

MCP Support

Local Model Specs

Quantization

FP8 mixed

Architecture

MoE (49B active)

Details

Release Date: April 24, 2026
Knowledge Cutoff: -
Source: Local
License: MIT
Model ID: deepseek-v4-pro

Last updated: May 27, 2026