DeepSeek V4 Photo: VCG

　　Chinese artificial intelligence (AI) company DeepSeek on Friday officially released its next-generation large language model, the DeepSeek-V4 Preview, which highlights a massive 1-million-token context window and formidable performance, while remaining open-source and free to use.

　　Two versions have been introduced this time: DeepSeek-V4-Pro, whose performance benchmarks against top-tier closed-source models like GPT, Claude, and Gemini, making it suitable for complex reasoning, coding, and building intelligent agents; and DeepSeek-V4-Flash, which is faster and more cost-effective, performing similarly to the Pro version on simple tasks but slightly weaker on complex ones, according to the company"s official WeChat account.

　　From now on, a 1M context window will be the standard configuration for all of DeepSeek"s official services, said the WeChat post. It means the latest model can read documents, novels, codebases, or reports equivalent to one million context units in one go, while being more memory-efficient and computationally lean than previous models.

　　The new model possesses rich world knowledge, significantly leading other open-source models in tests and trailing only slightly behind the top-tier closed-source model Gemini-Pro-3.1. In evaluations for Mathematics, STEM, and competitive-level coding, DeepSeek-V4-Pro has surpassed all currently documented open-source models, achieving outstanding results on par with the world"s leading closed-source systems, said the company.

　　Beyond this, compared to its predecessors, the Agent capabilities of DeepSeek-V4-Pro have been significantly enhanced, reaching the best level among current open-source models according to evaluations, and performing equally well in other Agent-related benchmarks.