DeepSeek V4 Professional 0813 (on OpenRouter). The most recent DeepSeek Professional mannequin is now obtainable, by way of API solely. I needed to hyperlink to OpenRouter as a result of DeepSeek have no apparent announcement web page for his or her new mannequin.
I have not been in a position to verify in the event that they plan to launch the open weights, however given the weights can be found for each April’s deepseek-ai/DeepSeek-V4-Professional and July’s deepseek-ai/DeepSeek-V4-Flash-0731 it appears seemingly. Replace: the weights are actually obtainable on Hugging Face, 1.7T parameters, 893 GB.
Apparently I bought very completely different trying pelicans for the three completely different reasoning ranges of low, medium, and excessive. I’ve not observed this sort of distinction from every other mannequin:
Low:

Medium:

Excessive:

When it comes to benchmarks… so far as I can inform these have been launched to the Official DeepSeek WeChat Group, then copied and pasted right into a publish on Reddit which was deleted by the moderators for being “low-effort”, then copied into this ASCII-art desk on Hacker Information.
