DeepSeek V4 Pro 0813 (on OpenRouter)
DeepSeek V4 Pro 0813 launches with 1.7T open weights and reasoning-level-dependent outputs.
“I've not noticed this kind of difference from any other model”
DeepSeek quietly released V4 Pro 0813 via API with no official announcement page, but weights (1.7T parameters, 893 GB) are now on Hugging Face. A notable behavioral quirk: the model produces visually distinct outputs across low/medium/high reasoning levels, something observer Simon Willison says he hasn't seen in other models. Benchmarks only surfaced informally through WeChat and a since-deleted Reddit post, highlighting DeepSeek's unconventional release practices.