On April 3rd, according to Late PostDeepSeek THE NEXT-GENERATION FLAGSHIP MODEL V4 IS EXPECTED TO BE RELEASED IN APRIL THIS YEAR. IN JANUARY OF THIS YEAR, THE V4 SMALL-PARAMETER VERSION WAS RELEASED TO SELECTED OPEN-SOURCE FRAMEWORK COMMUNITIES FOR ADAPTATION, AND THE LARGE-PARAMETRIC VERSION WAS ORIGINALLY SCHEDULED TO GO ON LINE AROUND SPRING BUT WAS EVENTUALLY POSTPONED。
Notably, the report states thatV4 IS PROBABLY STILL THE STRONGEST MODEL IN OPEN SOURCE, BUT IT'S HARD TO CRUSH THE PRESSURE LEVEL.
With the increasing diversity of AI assessment criteria, it has become difficult to measure the full range of model capabilities in Benchmark, especially in the context of the Agenic application, where the importance of hands-on and long-tailed data from products has become more and more important — which is precisely the area of DeepSeek's previous limited input。
The report also refers to DeepSeek's "anti-infiltration" culture: most employees leave the company at 6-7 p.m. on Sunday, without a card in the morning and without clear performance appraisals and deadlines。
The founder, Liang Wenbing, has a direct argument: "It is very difficult for a person to work more than 6-8 hours a day, while overworked judgement is a waste of valuable computing resources. I'm sorry