DeepSeek Web Version Quietly Launches "Fast" and "Expert" Dual Modes, Introducing Product Layering for the First Time

According to 1M AI News monitoring, DeepSeek’s web interface has added a mode-switching feature. Two icons—a lightning bolt and a diamond—appear above the input box, corresponding to “Quick Mode” and “Expert Mode,” respectively. Quick Mode is suitable for everyday conversations, delivers responses instantly, and supports image and file recognition. Expert Mode is aimed at complex reasoning tasks, but currently does not support file uploads or multimodal capabilities. The update went live completely silently, with no official announcement.

Community analysis shows that behind Quick Mode is DeepSeek 3.2, with a knowledge cutoff of July 2024. Expert Mode points to a newer model, possibly an early version of V4. In user tests, Expert Mode performs significantly better on deep reasoning tasks such as physical simulation and mathematical reasoning, but the gap versus Quick Mode is limited on simpler tasks like creative writing. Some testers believe that Expert Mode’s current routing is still to some version of V4 Lite, and that the full version of V4 will require more waiting. In the front-end code, a third option that has not yet been launched—“Vision Mode”—was also found. However, reverse engineering indicates that it is not an independent model; instead, a visual understanding parameter is enabled within Quick Mode.

This is DeepSeek’s first time doing product tiering since it went viral at the beginning of last year. Previously, the web interface was completely free and had no functional distinctions; it would route users based on their needs to different model entry points, which in itself is a compute-scheduling strategy. And once this architecture is in place, technically there is no longer any barrier to introducing a paid system later or imposing mode-based limits.