FLOPs2.0001e+25

Notes: Based on peer models such as Kimi K2.6 and DeepSeek V4 Pro, plausibly was trained on roughly 30T tokens (likely between 15T and 60T). Using the heuristic that C = 6ND, 30T pretraining tokens would imply ~1.9e25 FLOP. This is a somewhat uncertain estimate (more uncertain than estimates based on disclosed active params and token counts, but comparable to estimates based on disclosed GPU-hours). K3 was likely pretrained on over 1e25 FLOP. Speculatively, K3's relatively large size compared to other Chinese models may have reduced the training token budget vs smaller models.