DeepSeek touts model that outdoes flagship in agentic AI step

[BEIJING] DeepSeek unveiled an update to an older model that it says surpasses the seminal R1 on key benchmarks, keeping the Chinese startup in the game while the industry awaits its next flagship offering.

The V3.1 returns answers to queries much faster and marks the startup’s first step towards creating an AI agent, DeepSeek said in a WeChat post on Thursday (Aug 21). DeepSeek first outlined the V3.1 earlier this week, but the platform only just made it to the Hugging Face portal. The version has been customised to work with next-generation Chinese-made AI chips, DeepSeek said in a separate message.

DeepSeek shook Silicon Valley in January with the powerful yet low-cost R1 and has since remained at the vanguard of a growing wave of Chinese AI development. On Thursday, it also outlined a new pricing plan for its update to the longer-established V3, raising prices in some cases and abolishing discounted evening rates. It did however cut fees for certain use cases. The changes take effect on Sep 6.

DeepSeek’s models have challenged US incumbents such as OpenAI and demonstrated how Chinese companies can make strides in artificial intelligence without the most cutting-edge semiconductors.

Industry watchers await the release of the successor to R1, which had been expected to emerge earlier this year. Local media have attributed the delay to founder Liang Wenfeng’s determination to get it right, even as he continues to run his lucrative High-Flyer Asset Management outfit. Others have speculated about various glitches in training or development.

At the same time, rivals from Alibaba Group Holding to Tencent Holdings have kept up a frenetic pace of AI model development and updates. Alibaba’s Qwen models have, in particular, gained a popular following. BLOOMBERG