DeepSeek V4 Pro
All models
DeepSeekDeepSeekDeepSeekReleased April 24, 2026

DeepSeek V4 Pro

1.0M context$1.74/M input$3.48/M output1.6T total, 49B activated

DeepSeek V4 Pro is a large-scale Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion activated parameters. It supports a 1 million-token context window and is designed for advanced reasoning, coding, and long-horizon agent workflows. The model introduces a hybrid attention system for efficient long-context processing and supports multiple reasoning modes to balance speed and depth depending on the task, making it suitable for complex workloads such as full-codebase analysis and large-scale information synthesis.

What is DeepSeek V4 Pro?

DeepSeek V4 Pro is an AI model from DeepSeek that Agent Mag tracks for pricing, context window, modalities, benchmarks, and API compatibility. Builders can use this page to compare DeepSeek V4 Pro against other models for agent workflows and production deployments.

Model ID

DeepSeek V4 Pro is a large-scale Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion activated parameters. It supports a 1 million-token context window and is designed for advanced reasoning, coding, and long-horizon agent workflows. The model introduces a hybrid attention system for efficient long-context processing and supports multiple reasoning modes to balance speed and depth depending on the task, making it suitable for complex workloads such as full-codebase analysis and large-scale information synthesis.

Architecture & Specifications
Architecture
Mixture of Experts (MoE)
Parameters
1.6T total, 49B activated
Tokenizer
DeepSeek
Released
April 24, 2026
Modalities
Input
text
Output
text
Supported Parameters
frequency_penaltyinclude_reasoninglogprobsmax_tokenspresence_penaltyreasoningresponse_formatseedstoptemperaturetool_choicetoolstop_ktop_logprobstop_p
Strengths
  • Supports a 1 million-token context window
  • Advanced reasoning capabilities
  • Efficient long-context processing with hybrid attention system
  • Multiple reasoning modes for task-specific optimization
  • Well-suited for complex workloads like full-codebase analysis and multi-step automation
Limitations
  • Lower performance on research-level physics reasoning (CritPt benchmark: 12.9%)
  • Moderate accuracy on general knowledge tasks (AA-Omniscience Accuracy: 43.3%)
  • Hallucination rate of 6.0% in knowledge tasks
  • Limited coding proficiency (SciCode benchmark: 50.0%)
  • Performance variability across benchmarks
Recommended Use Cases
Full-codebase analysis
Multi-step automation workflows
Large-scale information synthesis
Advanced reasoning tasks
Scientific computing and coding

Related content

Data enriched Apr 24, 2026. Pricing from OpenRouter API.