DeepSeek released a preview version of its new V4 model series and made it open source. The model supports an ultra-long context of up to one million tokens and shows improvements in agent capabilities, knowledge coverage, and reasoning performance. The V4 series comes in two versions: Pro and Flash. The Pro version targets high-performance use cases, with overall capabilities approaching those of some leading proprietary models. The Flash version focuses more on efficiency and cost control, making it suitable for scenarios with tighter compute and budget constraints.
The models are currently available via DeepSeek’s official website and app. The API has also been updated to support compatibility with OpenAI and Anthropic interface standards. [TechNode Reporting]