DeepSeek V4 may arrive this month; test UI hints at Vision and Expert modes

New version could land soon

DeepSeek (深寻) — one of the more closely watched entrants in China’s domestic large model race — reportedly plans to launch DeepSeek V4 this month, according to a test interface seen by observers. The leak suggests two distinct modes in the new build, labelled “Vision” and “Expert,” marking a potential shift toward multimodal capabilities and domain-specialist behavior. It has been reported that the company opened a public test in February, and users have been awaiting an official release ever since.

What the interface reveals

The test UI points to an explicit Vision mode, implying integrated image understanding or generation alongside text — a table-stakes capability as Chinese models chase Western multimodal competitors. An “Expert” mode suggests a separate instruction set or tuning for professional tasks, hinting at more controlled, higher-precision outputs for vertical use cases such as finance, healthcare or legal work. It has been reported that the rollout timeline may have been accelerated after a recent surge in system crashes and anomalies in the test environment, though those operational issues remain unverified.

Market and geopolitical context

China’s large-model market is crowded and politically charged. Domestic players from Baidu (百度) to Alibaba (阿里巴巴) and many startups compete for enterprise customers while navigating stricter data and algorithm oversight at home. Geopolitically, U.S. export controls on advanced AI chips and rising trade tensions continue to shape how Chinese firms train and deploy large models, particularly for compute-intensive multimodal systems. Will DeepSeek V4 meaningfully shift competitive dynamics? Much depends on official confirmation, robustness in real-world use, and how regulators view increasingly capable multimodal AI.