- DeepSeek adds "expert mode," sparking speculation that the underlying model is the upcoming V4 release.
- The update comes ahead of the highly anticipated V4 launch, while a recent 12-hour system outage highlights ongoing compute bottlenecks.
Chinese AI startup DeepSeek has finally updated its features after a long period of silence, reigniting anticipation for its highly awaited V4 model.
DeepSeek's web interface added two new portals, "quick mode" and "expert mode," on Wednesday, marking the platform's first introduction of a tiered design and its first expert mode.
The quick mode is primarily designed for daily conversations with instant responses, and currently supports basic functions such as text recognition in images and files.
The expert mode handles complex problems and supports deep thinking, demonstrating a capability advantage in highly demanding tasks like mathematical and logical reasoning.
Currently, the expert mode does not support file uploads or multimodal functions, and requires queuing during peak hours.
The new update has triggered widespread speculation that the mode is likely running a lightweight test version of the upcoming V4 large model in the background.
DeepSeek's product tiering reflects the company's considerations regarding commercialization exploration and compute scheduling.
By diverting users with different needs into two separate portals, the platform can effectively limit quotas and traffic, thereby alleviating high system loads caused by compute shortages.
Just in late March, DeepSeek experienced a system crash lasting nearly 12 hours, setting a record for its longest single service interruption.
Against the backdrop of surging global demand for AI compute, balancing inclusive free services with high inference costs has become a challenge.
The launch of the expert mode may be a testing ground for DeepSeek as it gradually moves away from a completely free and untiered operating model.
