DeepSeek Marks Push Toward Agentic AI With New Model

DeepSeek said its new artificial intelligence model released this week surpasses its flagship R1 tool in some ways.

The launch of the V3.1 helps keep the Chinese company in the AI game as it prepares its latest iteration of its flagship model, Bloomberg reported Thursday (Aug. 21).

V3.1 returns answers to queries faster and is the first step to creating an AI agent, the company said, per the report. The version has been customized to work with next-generation Chinese-made AI chips.

DeepSeek rocked the tech world earlier this year when it debuted R1, showing how Chinese companies could take on high-profile rivals in the United States like OpenAI without the most cutting-edge semiconductors. The company’s family of models includes V3 (AI chat) and R1 (reasoning).

DeepSeek claimed to have spent just $5.6 million to train one of its model, less than the $100 million to $1 billion figure cited by Anthropic.

The successor to R1 was expected earlier this year, the Bloomberg report said. Chinese media attributed the hold up to DeepSeek’s determination to get the product right, while others speculated about training or development-related hiccups.

Meanwhile, a gap exists between awareness of agentic AI and interest in using it.

The PYMNTS Intelligence report “The Two Faces of AI: Gen AI’s Triumph Meets Agentic AI’s Caution” found that while nearly all finance chiefs surveyed said they were aware of the technology, just 15% showed any desire to deploy it within their companies.

The gap underscored lingering skepticism among business leaders about the maturity and business value of AI agents as they exist now. Although the technology is promising when it comes to automating complex workflows and improving decision-making, many CFOs remain hesitant due to concerns about implementation risks, oversight challenges and a lack of evidence of ROI.

“A lot of companies are excited about what agentic AI can do, but not enough are thinking about what it takes to use it safely,” James Prolizo, chief information security officer at Sovos, told PYMNTS this month. “These tools are starting to make real decisions, not just automate tasks, and that changes the game.”

The research also found that building trust depends on the ability to provide user-friendly reports and visualizations that clearly explain the reasons for an AI agent’s actions, as well as ways to provide ongoing human supervision and intervention for critical decisions.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

The post DeepSeek Marks Push Toward Agentic AI With New Model appeared first on PYMNTS.com.