DeepSeek V4 could be released this month, with the test interface indicating the presence of Vision and Expert modes.

      Credit: DeepSeek Logo

      In China's large model sector, the DeepSeek V4 is certainly among the most awaited releases in recent months. Since its test version launched in February, users have been eagerly anticipating the official version. With a recent increase in system crashes and irregularities, there's speculation in the market that these may serve as stress tests ahead of the new version's debut, which is expected to occur as early as this month. A recently leaked screenshot from a gray-scale test has heightened expectations for significant upgrades.

      A well-known programmer on the Chinese social media platform Weibo, @Yigongchang, shared a screenshot on Tuesday showcasing DeepSeek’s updated interface. Unlike the current version, which offers the familiar dual options of Deep Thinking (R1) and Smart Search (online), the new interface presents three additional modes: Fast, Expert, and Vision. This change indicates a redesign of the interaction logic and suggests a more diversified product range. Analysts suspect that the Fast mode may refer to a lightweight model, while Expert is likely associated with a deep-reasoning model. The prominent addition of a Vision option further suggests that the previously mentioned multimodal capabilities have now been confirmed, marking a significant step for DeepSeek in bridging the gap with leading large model competitors.

      DeepSeek V3.2

      In contrast to the current version, the forthcoming release of DeepSeek might not consist of a single model but rather a collection of offerings. Given the interface options, the new versions could be designated as DeepSeek V4 Lite (Fast version), DeepSeek V4 (Expert version), and DeepSeek V4 Vision (Vision version). Previous reports have indicated that DeepSeek is working on at least three large models and constructing them on a domestic AI chip-based computing platform, aligning well with the information from this gray-scale test.

      Although DeepSeek has not disclosed specific details about the upgrades, the launch of the V4 series appears to be just around the corner. Reflecting on the previous year, the introduction of DeepSeek R1 made waves in the market due to its impressive cost-performance ratio, even momentarily impacting NVIDIA’s market value.

      Credit: DeepSeek Logo

      Now, with the V4 series achieving its multimodal capabilities, there are high expectations domestically. Some in the industry are even advocating for a specialized version focused on AI programming, aiming to compete directly with Anthropic’s Claude series in the coding realm. Whether DeepSeek can once again harness its technological prowess to disrupt the market and challenge Anthropic or OpenAI remains to be seen—answers may soon emerge.

      Jessie Wu is a tech reporter based in Shanghai. She covers consumer electronics, semiconductors, and the gaming industry for TechNode. You can connect with her via e-mail: jessie.wu@technode.com.

      More by Jessie Wu