[Salon] Fwd: SCMP: "DeepSeek’s upgraded AI model absorbs reasoning feature in move towards ‘agent era’." (8/21/25.)

[Chas Freeman <cwfresidence@gmail.com>]

https://www.scmp.com/tech/tech-trends/article/3322669/deepseeks-upgraded-ai-model-absorbs-reasoning-feature-move-towards-agent-era?module=top_story&pgtype=homepage

DeepSeek’s upgraded AI model absorbs reasoning feature in move towards ‘agent era’

21 Aug 2025

Chinese artificial intelligence start-up DeepSeek said on Thursday that its newly released V3.1 model supported both “think” and “non-think” modes, marking the firm’s “first step towards the agent era” – a shift that suggests a change in its research focus and and the possibility it would forgo the highly anticipated R2 reasoning model.

The “think” mode on DeepSeek’s namesake chatbot was previously powered by its R1 reasoning model that garnered global attention after its release in January, following the launch of the V3 foundational model in December.

In contrast, the V3.1 model unveiled on Wednesday adopted a “one model, two modes” approach, indicating that the company may not develop a successor to R1.

The V3.1 model could deliver answers more quickly than R1, which was last updated in May, DeepSeek said on its official X account.

Founded by entrepreneur Liang Wenfeng as a side project of his quantitative trading firm, DeepSeek has spurred a wave of open-source AI adoption in China. The privately held company, however, has not disclosed its development timeline or future plans.

10:41

How Hangzhou’s ‘Six Little Dragons’ built a new Chinese tech hub

How Hangzhou’s ‘Six Little Dragons’ built a new Chinese tech hub

DeepSeek’s announcement comes as the start-up has been losing users in recent months, as open-source models from larger Chinese tech companies, such as Alibaba Group Holding’s Qwen family, gain traction in the domestic and international AI markets. Alibaba owns the South China Morning Post.

DeepSeek said V3.1’s enhanced agent capabilities lay the groundwork for the model to support AI agents – software that helps users automate specific tasks. Several start-ups, such as Manus AI, have been gaining attention for their AI agents.

In a statement on its WeChat account, DeepSeek said V3.1 was able to expedite the thinking process by up to 50 per cent without compromising reasoning capabilities.

The move to combine reasoning and non-reasoning features into a single model reflects a broader trend among AI developers to provide a unified interface for users. OpenAI’s recently launched GPT-5 incorporates a deeper reasoning model and uses a real-time router to determine whether a prompt should be directed to that model.

Liang previously said his long-term goal was to achieve artificial general intelligence – commonly defined as AI that can understand, learn and apply knowledge across a wide range of tasks at a level comparable to that of a human.

With the upgrade, the company has increased the pricing of its application programming interface (API) services for developers. From September 6, the charge for a million token input of API requests will rise to 4 yuan (US$0.56), while the output price will increase to 12 yuan per million tokens.