DeepSeek Launches DeepSeek-V4 Series Models and Open-Sources Them, Marking the Era of Million-Context Accessibility

DeepSeek officially launched the preview version of its new DeepSeek-V4 series models and open-sourced them simultaneously, marking the era of million-context accessibility for large models. The model uses innovative attention mechanisms and DSA sparse attention technology to achieve world-leading million-word long-context processing capabilities while reducing computing and memory requirements, with two versions: Pro (1.6 trillion parameters) and Flash.

DeepSeek-V4-Pro has achieved the best performance among open-source models in Agentic Coding evaluation, with world knowledge and reasoning capabilities surpassing open-source models and matching top closed-source models. API services have been updated to support OpenAI/Anthropic interfaces and 1M context length; old API model names will be discontinued in July 2026.

The open-source package includes model weights and technical reports, which developers can access via Hugging Face and ModelScope, driving AI applications to expand into complex scenarios.