DeepSeek V4 Launch Sparks Mixed Reactions: Architecture Praised, but Gap to Frontier Models Remains

According to ME News, on April 24 (UTC+8), monitoring by Beating revealed that since the release of DeepSeek V4, developer communities in China and the U.S. have simultaneously demonstrated technical recognition and a shared consensus on performance gaps. In the U.S., Replit CEO Amjad Masad praised V4’s attention compression and improved long-context efficiency as genuine architectural innovations. Developers on Reddit and Hacker News responded positively to the open-source 1M context model and MIT license, though most adopted a “wait-and-see” approach. CFR researcher Chris McGuire noted that V4’s own report acknowledges being 3 to 6 months behind state-of-the-art models. In China, communities on V2EX and Zhihu focused on V4’s programming agent capabilities and its low-price strategy; early feedback suggested the Pro version is approaching the level of Claude Opus 4.6, yet still lags in complex, deep reasoning tasks. Significant attention has also been paid to its compatibility with Huawei Ascend. Prior to launch, months of repeated delays had generated considerable skepticism, but post-release focus has shifted to intensive testing. Both communities agree: V4 performs strongly in coding and long-context scenarios, but has yet to match the overall reasoning capabilities of leading U.S. closed-source models. (Source: BlockBeats)