DeepSeek-V4 Officially Launched: The Era of Million-Token Context for Everyone

DeepSeek has officially released the V4 model series, marking a new era where million-token context becomes accessible to everyone. This launch includes DeepSeek-V4-Pro and DeepSeek-V4-Flash — two editions built for different workloads.

Dual-Edition Strategy

DeepSeek-V4-Pro

The Pro edition features 1.6T total parameters and 49B active parameters, making it the flagship model for complex tasks. It reaches state-of-the-art levels among open-source models in Agent coding, mathematical reasoning, and world knowledge.

DeepSeek-V4-Flash

The Flash edition offers 284B total parameters and 13B active parameters, delivering strong capability with lower latency and cost. It is ideal for everyday conversation, lightweight code generation, and other high-frequency use cases.

Key Breakthroughs

1. Million-token context: A 1M token context window is now standard across all official services
2. Agent capabilities: Agentic coding has improved significantly, with optimizations for leading Agent products such as Claude Code
3. Reasoning performance: MMLU-Pro 87.5%, GPQA Diamond 90.1%, and top-tier results on Chinese benchmarks
4. Open-source friendly: MIT license enables free community deployment

How to Try It

You can experience DeepSeek-V4 through:

• The web app at chat.deepseek.com
• The official mobile app
• The API, using model names deepseek-v4-pro or deepseek-v4-flash

Note: Legacy model names deepseek-chat and deepseek-reasoner will be retired on July 24, 2026. Please migrate as soon as possible.

The launch of DeepSeek-V4 brings ultra-long context, top-tier reasoning, and Agent capabilities to a much wider audience. Whether you are a developer, researcher, or everyday user, there is something here for you.