Skip to content

DeepSeek V4 Pro

deepseek-ai/DeepSeek-V4-Pro
Open Source · chat · open-weights · seen 7m ago
GA Alert me on changes
Context
1M
Max output
Input $/1M
$0.43
Output $/1M
$0.87
Modalities
text
Released
22 Apr 2026
License: mit · deepseek-ai/DeepSeek-V4-Pro
AI summary
● machine-written

DeepSeek V4 Pro: 1.6T-parameter Mixture-of-Experts model with 1M token context

DeepSeek V4 Pro is a large-scale Mixture-of-Experts model with 1.6T total parameters and 49B activated parameters, supporting a 1M-token context window. It uses hybrid attention architecture for efficient long-context processing. The model is designed for advanced reasoning, coding, and long-horizon agent workflows, with documented performance across knowledge, math, and software engineering benchmarks.

What's new
  • Hybrid attention system combining CSA and HCA for long-context efficiency
  • 27% of single-token inference FLOPs compared to V3.2 at 1M token prompts
  • Reasoning efforts high and xhigh supported; xhigh maps to max reasoning
  • Available through multiple inference providers including Together, Novita, Fireworks AI, and others
Best for
Advanced reasoning and complex reasoning tasksFull-codebase analysis and software engineeringLong-horizon agent workflows and multi-step automationLarge-scale information synthesis
Sources

Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro