$2.6M RL run nears SOTA, sparking training-cost debate

Xiaomi's MiMo results—near-SOTA performance from just 130 hours, 75B tokens and $2.6M of RL training—sparked debate over commoditized model training. Commenters also questioned compute accounting and whether data costs could eventually exceed training costs.

2026-09-21 ~ 2026-09-23 · 4 related posts

• MiMo near-SOTA on DeepSWE with just ~$2.6M RL run: will data cost more than training? — my_cat_can_code · 2026-09-21
• One RL Run Cost 130 Hours, 75B Tokens and $2.6M — the Real Price of Scaling RL — burny_tech · 2026-09-23
• Xiaomi's $2.6M RL run: 130 hours, 75B tokens fuel 'commoditize training' debate — burny_tech · 2026-09-23
• Frontier RL compute is shockingly cheap: 50 scale units prompt questions about what labs actually spend — teortaxesTex · 2026-09-23