Actually we tried Cerebras. It caps output at 32K tokens so the 128K reasoning (the author budget) didn't fit. Cerebras speed is real though
Hacker News new | past | comments | ask | show | jobs | submit