On April 24th, Beijing Time, DeepSeek officially launched its new generation large model, DeepSeek-V4. Its overall performance has reached a level comparable to current top closed-source models, becoming another milestone product in the open-source AI field.
Ultra-long 1 million token context window
DeepSeek-V4 supports an ultra-long context of up to 1 million tokens, providing a solid foundation for complex long document analysis, ultra-long code generation, and multi-turn Agent tasks.
Two versions strategy, flexible to meet different scenarios
The new model comes in two versions:
- DeepSeek-V4-Flash: lightweight and efficient version, suitable for fast iteration and daily development
- DeepSeek-V4-Pro: high-performance flagship version, tailored for extreme demand scenarios
Comprehensive excellence in mathematics, STEM, and competition code
In mathematics, STEM, and competition code evaluations, DeepSeek-V4-Pro surpasses all currently publicly evaluated open-source models, demonstrating overwhelming strength.
Significant improvement in Agent capabilities, reaching new heights in open-source
The Agent capabilities of DeepSeek-V4-Pro have been significantly enhanced. In Agentic Coding specialized evaluations, it has reached the best level among current open-source models and performs excellently in multiple Agent-related evaluations, paving the way for the implementation of complex autonomous tasks.
Leading world knowledge reserves, second only to top closed-source models
In world knowledge assessments, DeepSeek-V4-Pro greatly outperforms other open-source models, trailing only slightly behind the top closed-source model Gemini-Pro-3.1, achieving a qualitative leap in knowledge breadth and depth.
Pricing that shatters the market, making the cost of 1 million tokens reach rock bottom
DeepSeek-V4 continues its series'一贯 extremely cost-effective approach:
- Input: 1 million tokens for Flash version cost 1 yuan, and 12 yuan for Pro version
- Output: 1 million tokens for Flash version cost 2 yuan, and 24 yuan for Pro version
This pricing strategy once again breaks industry boundaries, bringing the capabilities of top models into the hands of ordinary developers and small and medium-sized enterprises.
AIbase will continue to track the practical application performance and community feedback of DeepSeek-V4, please stay tuned for subsequent in-depth evaluation reports.
