The official version of DeepSeek V4 has been tested and exposed, and it may be released next Monday to target Kimi K3. As the global AI circle focuses on the popularity of Kimi K3, there are new trends from the domestic large model manufacturer DeepSeek

The official version of DeepSeek V4 has been tested and exposed, and it may be released next Monday to target Kimi K3. As the global AI circle focuses on the popularity of Kimi K3, there is a new trend from the domestic large model manufacturer DeepSeek. Recently, the test video and actual machine demonstration of the official version of DeepSeek V4 were quietly exposed on social platforms. Multiple sources pointed out that the model is expected to be officially released as early as next Monday (the first Monday in late July). The timing coincides with the launch of K3, which has attracted great attention in the industry. According to leaked test information and feedback from overseas bloggers, the official version of DeepSeek V4 performs amazingly on a number of core indicators. The measured performance has initially reached the level of Claude Opus4.8, especially the generation capability in the 3D direction far exceeds Opus4.8. Agent behavior and coding robustness have also been significantly improved. A more radical view is that in specific tests (such as game code generation), the official version of V4 has surpassed Fable5 and GPT-5.6. In a coding test of the "Squid Game" mini-game shown by a blogger, the total cost of generating 3,000 lines of code with V4 was only US$0.12, highlighting its incredible cost-effective advantage. In addition to improvements in hard performance, V4 also faces fine-tuning in the interactive experience. Testers noticed that its chain of thought (CoT) style is closer to Claude. It no longer piles up code walls, but is presented with summary steps such as "writing core functions" and "writing main scene codes", which greatly enhances readability. Some analysts believe that V4 may have refined Fable's coding style and made the output look and feel more concise. Looking back at the past rhythm of DeepSeek, we can see that it has always used the strategy of "working hard and butchering the price" to break through. If V4 can really approach the current top model level at a price one order of magnitude lower, it will definitely set off the "DeepSeek Moment 2.0" after V3. Currently, the official announcement is that it will be released in mid-July. The timeline for next Monday is in line with expectations. The final release still needs to wait for the official announcement. via AI News (author: AI Base)