小米披露MiMo-V2.6规模化强化学习路线：单次RL后训练耗资260万美元，每步处理37亿Token。
The $2.6M cost per RL run is a concrete, source-verified figure — not an estimate — revealing the staggering infrastructure burden behind production-grade AI agent alignment. This applies specifically to post-training behavioral tuning, not base model training, making it highly relevant for startups building agentic workflows. For early teams, this signals that 'RLHF-like' refinement at scale is financially prohibitive without strategic shortcuts: e.g., synthetic preference data, reward model distillation, or constrained action spaces. Teams must treat compute budget as a first-class product requirement — not an afterthought — and prioritize architectures where behavior emerges from lightweight, composable modules rather than monolithic, billion-token RL loops.