DeepSeek-V4: Detailed Introduction

[ikaijua]

DeepSeek-V4: Detailed Introduction

Table of Contents

• What is DeepSeek-V4?
• Core Capabilities
• Access
• Pricing

What is DeepSeek-V4?

DeepSeek-V4 is DeepSeek's fourth-generation flagship model, released as open weights under the MIT license on Hugging Face. It includes V4-Pro (1.6T MoE, 49B activated) and V4-Flash (284B MoE, 13B activated), both supporting a 1-million-token context and native multimodality.

Core Capabilities

• Mixture-of-Experts architecture: 1.6T total parameters with sparse activation for efficient inference.
• Long context: 1M-token context window for long-document and repository-scale tasks.
• Native multimodality: handles text, image, and video understanding and generation.
• Hybrid Attention (CSA + HCA): reduces long-context inference FLOPs and KV-cache usage.
• Three reasoning modes: Non-think (fast), Think High, and Think Max; Pro-Max leads on coding and agentic benchmarks.

Access

Resource	Link
Hugging Face collection	https://huggingface.co/collections/deepseek-ai/deepseek-v4
DeepSeek chat	https://chat.deepseek.com
DeepSeek API	https://platform.deepseek.com

Pricing

The model weights are free under MIT. API access is usage-based; V4-Flash offers a lower-cost option while V4-Pro delivers top-tier performance.