On April 8th, eagle-eyed users of DeepSeek’s web version may have spotted a subtle yet significant change. Adjacent to the input box, two fresh options made their debut: ‘Fast Mode’ and ‘Expert Mode.’

Let’s delve into ‘Fast Mode’ first, whose purpose is straightforward. It essentially maintains the original DeepSeek experience, adept at handling daily queries and facilitating instant conversations with a strong emphasis on efficiency. However, it’s the ‘Expert Mode’ that has truly piqued the industry’s interest. This mode transcends mere question-answering, striving instead to offer comprehensive problem-solving solutions.

From practical usage, Expert Mode delivers markedly more professional and in-depth responses in specialized fields such as programming, law, and medicine, steering clear of generic replies. Crucially, when tackling complex, multi-step reasoning tasks, it exhibits a capacity akin to human-like ‘thinking.’ This entails more than just expanding the knowledge base; it involves extending and reinforcing a coherent chain of thought.

The distinction is perhaps best exemplified by a classic brain teaser: ‘Can a 7-meter-long sugarcane pass through a door measuring 2 meters in height and 1 meter in width?’

‘Fast Mode’ provides an intuitive, quintessentially human response: No. This aligns with the initial reaction of most individuals when confronted with the physical constraints of space. Conversely, ‘Expert Mode’ offers a counterintuitive yet accurate answer: Yes. It elucidates that the sugarcane’s small cross-sectional area enables it to be tilted and passed through the diagonal space of the doorframe, much like a spear. It doesn’t merely ascertain the ‘whether’ but also elucidates the ‘how.’

This simple test perhaps underscores the essence of this upgrade: transitioning from providing ‘informational answers’ to offering ‘solutions,’ from ‘answering what’ to ‘explaining why.’

This notable enhancement in capability naturally prompts the industry’s most pressing query: Has the underlying model been altered? Indeed, numerous industry insiders speculate that ‘Expert Mode’ is likely underpinned by a larger, more potent model—potentially a version of the much-anticipated DeepSeek-V4 official release. Sources suggest that V4 will be rolled out in mid-to-late April. Feel free to share your insights in the comments section.

Compared to its predecessors, Expert Mode demonstrates enhanced robustness and depth in mathematical reasoning, logical chain derivation, and complex problem dissection. It’s not merely about ‘mimicking human-like responses’; it’s about striving to be ‘more reliable than the average person’ in domains demanding professionalism and rigor.

However, if you’re anticipating an ‘omniscient expert,’ you may need to exercise a bit of patience. The currently available Expert Mode exhibits a certain degree of ‘restraint.’ It does not yet support file uploads or multimodal recognition capabilities. This implies that its ‘expertise’ is currently focused on in-depth text processing and reasoning.

This phased approach to capability release can perhaps be interpreted as DeepSeek’s pragmatic strategy: first deepening, refining, and stabilizing the core capability of textual logical reasoning before venturing into other areas. For professional users who genuinely require the processing of lengthy documents, complex analytical reports, and intricate code, a robust textual reasoning core may be more fundamental than flashy, superficial features.

In today’s context, this might represent the ‘long-termism’ of the AI field—eschewing short-term gimmicks in favor of silently ‘honing internal skills’ in the realm of professionalism and depth. Because true ‘intelligence’ ultimately resides in its ability to genuinely ‘solve problems’ for individuals.