DeepSeek has introduced an upgrade to its AI model that merges both “think” and “non-think” modes, a step the company describes as its “first step towards the agent era.” According to South China Morning Post, the newly released V3.1 model lays the groundwork for supporting AI agents.
V3.1 blends reasoning and non-reasoning modes
The update, identified as V3.1, incorporates both reasoning and non-reasoning capabilities. As reported by SCMP, DeepSeek described this integration as an architectural move that enables the model to operate in “think” and “non-think” modes. The company framed the release as a milestone toward AI agents, highlighting a change in research direction.
SCMP noted that the shift also suggests the possibility that DeepSeek could forgo its highly anticipated R2 reasoning model. The company’s positioning of V3.1 indicates a focus on agentic capabilities, with the combined modes intended to support tasks aligned with that approach.
Positioning toward the ‘agent era’
DeepSeek characterized the V3.1 update as a foundational step toward agent-oriented systems. By bringing reasoning features into the core model alongside non-reasoning operations, the company signaled a broader framework aimed at facilitating agent behaviors, according to the report.
Implications for DeepSeek’s roadmap
The move comes amid attention on DeepSeek’s model lineup and the expectations surrounding future releases. As described by South China Morning Post, the emphasis on V3.1’s dual modes underscores the company’s evolving research priorities. The report linked the update to a potential shift away from a standalone R2 reasoning model in favor of integrated functionality within V3.1.
In outlining V3.1, DeepSeek presented the upgrade as a platform for agent-supportive capabilities. The framing of the release as an initial step toward the “agent era” reflects the company’s intent to align model features with agent-centric use cases. SCMP’s coverage emphasizes the combined “think” and “non-think” operation as central to that trajectory.
The company’s announcement positions V3.1 as a key update that merges reasoning with non-reasoning functions. According to the SCMP report, this alignment is meant to prepare the model for agent applications and signals the direction of DeepSeek’s ongoing development.
