Gen-1 Slides: Opus 5-level decks at a fraction of the cost

This post details the post-training of MiniMax M3 into Gen-1 Slides using reinforcement learning, a process managed by Fireworks Lab in partnership with Genspark. Key technical contributions include: encoding a complex evaluation standard for deck quality, designing and executing a curriculum-based RL training approach (SFT, short-context RL, then extended context), and implementing systems-level solutions to maintain numerical stability for long-horizon trajectories (100,000+ tokens). Specific challenges addressed include tokenization mismatches between training and inference, and credit assignment across dozens of turns. The outcome is a model matching Opus 5 quality at a fraction of the cost, with a significant reduction in low-rated outputs.