DeepSeek Unveils Flagship AI 'V4': "Top-Tier Performance in Coding Benchmarks"

Chinese artificial intelligence (AI) company DeepSeek has unveiled its latest AI model, the 'V4.'

Reuters Yonhap News

On April 24, DeepSeek announced via the social networking service X (formerly Twitter) that it had officially released the DeepSeek-V4 preview as open source, introducing two models: the open-source-based V4 Flash and V4 Pro.

DeepSeek explained that this model has achieved top-tier performance in coding benchmarks and made significant advancements in reasoning and agent-based tasks. In particular, the company highlighted its technology named 'Hybrid Attention Architecture.' This technology is said to enhance the AI platform's ability to retain queries throughout extended conversations. Additionally, DeepSeek expanded the context window to one million tokens, which enables entire codebases or lengthy documents to be input as a single prompt.

DeepSeek first introduced its large language model (LLM) 'V3' in December 2024. In January 2025, it released the low-cost model 'R1.' The newly released V4 comes roughly one year after the launch of R1. R1 was recognized for achieving performance close to major U.S. commercial models at a much lower development cost than its American counterparts.

According to benchmark results released by DeepSeek on the same day, the V4-Pro outperformed major open-source models and demonstrated competitive performance even against some of the highest-ranking closed-source models.

Nonetheless, DeepSeek emphasized that its focus is not solely on maximizing model performance but also on fundamentally reducing costs. V4 is designed to operate on more affordable infrastructure. Beison Ling, Managing Director at Union Bancaire Privee, told Bloomberg News, "Independent model providers such as MiniMax and Zhipu are structurally vulnerable in terms of reach and distribution compared to internet platforms or cloud service providers, which have a stronger advantage in these areas. Ultimately, most users will barely notice the performance gap between the models."

This content was produced with the assistance of AI translation services.