DeepSeek V4, which was more than two months late, has finally arrived. If it weren't for the adaptation of Chinese chips, it would have been absolutely explosive if it had been released two months ago. Although it is outstanding now, it has not yet reached a stunning level. I have finished reading the technical report on page 58 of DeepSeek, and the conclusion is placed at the beginning: This generation V4 is not just a minor fix, but an upgrade of the open-source model camp. 1.6T total parameters, 49B activation, native 1M token context, single token inference FLOPs reduced by 73% compared to V3.2, KV cache reduced by 90%. In the comparison table provided by themselves, Kimi 2.6 (K2.6) and GLM 5.1 have a 10-20 percentage point gap in most knowledge and reasoning tasks, while Codeforces directly scored 3206 points (ranked 23rd among humans). DeepSeek's open source competitors this time are K2.6 Thinking and GLM-5.1 Thinking, which are about the same strength as the Qwen 3.6 max preview version, but DeepSeek's pricing is lower, which makes Qwen embarrassed. The ranking has not been stable for two days yet.