In China’s domestic large model sector, DeepSeek V4 is undoubtedly one of the most highly anticipated releases in recent months. Since the launch of its test version in February, users have been eagerly awaiting the official release.
Coupled with a recent surge in system crashes and anomalies, the market speculates that these may be stress tests ahead of the new version’s rollout, with a full launch expected as early as this month. A recently surfaced screenshot from a gray-scale test has further fueled expectations of a major upgrade.
A well-known programmer on Chinese social platform Weibo, @Yigongchang, on Tuesday shared a screenshot of DeepSeek’s updated interface. Unlike the current version, which features the familiar dual options of Deep Thinking (R1) and Smart Search (online), the new interface clearly shows three additional modes: Fast, Expert, and Vision.
This shift signals a redesign of the interaction logic and points directly to a more segmented product lineup. Analysts believe the Fast mode may correspond to a lightweight model, while Expert is likely tied to a deep-reasoning model.
The prominent inclusion of a Vision option further indicates that the previously reportedly multimodal capabilities are now confirmed, marking a crucial step for DeepSeek in closing the gap with top-tier large model competitors.
Compared with the current version, DeepSeek’s upcoming release may not be a single model, but rather a suite of offerings. Based on the interface options, the new versions could be named DeepSeek V4 Lite (Fast version), DeepSeek V4 (Expert version), and DeepSeek V4 Vision (Vision version).
Previous reports have indicated that DeepSeek is developing at least three large models and is building them on a domestic AI chip-based computing platform, making the information from this gray-scale test highly consistent with earlier leaks.
Although DeepSeek has yet to reveal specific upgrade details, the release of the V4 series appears imminent. Looking back at last year, the launch of DeepSeek R1 shook the market because it had a high cost-performance ratio. It even briefly impacted NVIDIA’s market value.
Now, with the V4 series completing its multimodal capabilities, domestic expectations are high. Some industry voices are even calling for a special version aimed at AI programming, positioning it to directly compete with Anthropic’s Claude series in coding capabilities.
Whether DeepSeek can once again leverage its technological strength to shake up the market and put pressure on Anthropic or OpenAI remains to be seen—answers may emerge in the coming days.
