The 671b model should not have been used. Instead, a small distilled model was used.
Instead of speculating, give proof with prompts and responses as one-shots.
I don’t know. In my opinion, at least based on my very limited testing, it looks like a full-fat R1. Though there was one issue - Cursor seems to have a timeout, probably around 5 minutes, which, combined with slow inference speeds, can lead to terrible results (cut-off responses mid-thinking without any result).
Hey, we should have the cut off issue fixed, DeepSeek R1 was out slowest model to date when we first added it, but it’s now much quicker!