Chinese AI unicorn DeepSeek is keeping the industry guessing regarding the release of its next-generation models, V4 and R2. Despite intense speculation about a potential launch during the upcoming Lunar New Year, the company has declined to comment on specific dates.
Instead, the firm is letting its research do the talking. In recent weeks, DeepSeek has published multiple technical papers outlining novel methods to bypass current hardware limitations. With US sanctions restricting China’s access to advanced Nvidia GPUs, DeepSeek’s latest research focuses on algorithmic efficiency.
On Tuesday, the company introduced “Engram,” a conditional memory technique designed to solve GPU high-bandwidth memory shortages—a critical bottleneck in scaling AI. This follows a paper co-authored by CEO Liang Wenfeng on “manifold-constrained hyper-connections” (mHC), a framework aimed at training massive systems more cost-effectively.
Analysts believe these innovations are strategic moves to maintain momentum despite geopolitical headwinds. “DeepSeek just wants to prove that AI infrastructure innovation would drive efficiency,” said Zhang Ruiwang, a Beijing-based information systems architect.
While the release date for the successors to the V3 and R1 models remains a mystery, the company confirmed that its core team remains intact. An updated paper on the R1 model noted that all 18 original core scientists are still with the company, defying China’s high industry turnover rates.
