Slow inference is also present on z.ai, eyeballing it the 4.7 flash model was twice as slow as regular 4.7 right now.