DeepSeek has introduced an update to its V3 AI model, named V3-0324, which reportedly enhances programmability while setting new standards for accuracy and efficiency, Bloomberg reports. The update was quietly published on the Hugging Face platform without an official announcement.
The open-source DeepSeek V3 AI model was initially launched late last year. It utilizes a Mixture of Experts (MoE) architecture, featuring 671 billion total parameters and 37 billion parameters activated per token. At the time of release, DeepSeek stated that it spent $5.5 million training V3—significantly less than the amounts invested by competitors like OpenAI in similar models.
DeepSeek’s Competitive Edge
A few weeks after the V3 model’s debut, DeepSeek released R1, an open-source reasoning model. Despite its modest development budget, R1 outperformed OpenAI’s o1 model in multiple benchmarks. The rapid progress of the Chinese startup raised concerns among investors regarding the substantial costs associated with AI development in the U.S.
In January 2025, DeepSeek surpassed ChatGPT to become the most downloaded free app on the Apple App Store in the United States, reminds NIX Solutions. This success contributed to broader market concerns, leading to a collapse in the AI sector. Many companies experienced sharp declines in stock prices, with industry leader Nvidia losing $593 billion in market value in a single day—the largest one-day drop in stock market history.
We’ll keep you updated on further developments as DeepSeek continues to advance its AI models.