@MaxForAI: Breaking: DeepSeek-V4-Pro-0813 was indeed the wrong model uploaded. Yesterday, when the community inspected the V4-Pro-0813 that DeepSeek uploaded to Hugging Face, they found a very unusual problem: the open-source model architecture exactly matched the previous V4...

Breaking: DeepSeek-V4-Pro-0813 was indeed the wrong model uploaded Yesterday, when the community was inspecting the V4-Pro-0813 that DeepSeek uploaded to Hugging Face, they found a very unusual problem: The open-source model architecture actually matched the previous V4 Flash exactly, not the V4-Pro-preview. Specifically: hidden_size: 4096 43 layers 256 routed experts 64 attention heads These parameters are basically the same as V4-Flash. A normal V4 Pro should be: hidden_size: 7168 61 layers 384 routed experts 128 attention heads What's more suspicious is that after the community discovered the issue, the V4-Pro-0813 repository on Hugging Face was taken down for a while. Then, around 0:00 Beijing time today, DeepSeek re-uploaded the model and changed the config back to the Pro architecture. At this point, it could still be explained as: they just copied the wrong config.json when open-sourcing. But later, people continued to check the commit history on Hugging Face and found that after the re-upload, not only did the config change, but the SHA256 hashes and file sizes of some safetensors weight files also changed. In other words, DeepSeek likely didn't just modify a config file—they may have reprocessed, re-exported, or even quantized part of the checkpoint. This makes the various abnormal behaviors seen after the V4 Pro 0813 release yesterday even more worthy of re-examination. A 1.6T-class flagship model release may ultimately have tripped over an incorrect deployment. They managed to train the model, but not to ship it properly. The world really is one big makeshift operation.