Perf modeler revises K3 cost estimate after precision error: still ~20x pricier
z4y5f3 · x · 2026-10-26
z4y5f3 accepts YouJiacheng's FP8 KV cache numbers, admitting he had used HF default precision (BF16 for MLA, FP32 for KDA), which inflated the KV gap to 40x. He notes FP8 runs faster too and doesn't change the compute-vs-bandwidth equation, revising his conclusion: K3 is still 20x more expensive to serve than V4.1F, not 6x. Both sides agree K3 serving costs are substantially higher.