It's Time to Re-evaluate MiniMax

Genspark, typically an API-agnostic AI application company, has begun training its own model—a significant shift in strategy. On September 10, it released Gen-1 Slides, a specialized PPT-generation model built on MiniMax M3 in collaboration with Fireworks, now the default for its standard mode. This move signals that open-source models have reached sufficient maturity for application companies to invest in post-training rather than building from scratch.

MiniMax M3 emerged as the optimal choice not because it ranks highest on leaderboards, but because it strikes a practical balance. With approximately 428B parameters and 23B activated per token, M3 is large enough to handle complex tasks—multimodal understanding, long context, coding, and tool use—yet small enough that post-training remains economically viable for application companies. Larger models, while potentially more capable, impose prohibitive costs for reinforcement learning, deployment, and iteration. For application teams, the goal is not the smartest foundation model, but the best cost-performance ratio among capability, cost, and trainability.

The economics of custom training became compelling at Genspark's scale: over 1 trillion tokens consumed monthly. Testing revealed Gen-1 Slides reduces cost per PPT from $4.16 (Opus 5) to $0.44—a tenfold reduction—while maintaining quality (4.25 vs. 4.23 rating) and improving download rates (33.1% vs. 31.5%). This demonstrates that specialized fine-tuning for high-frequency, stable tasks can deliver both cost savings and competitive performance. The lesson extends beyond Genspark: as open models mature and token consumption scales, application companies increasingly find it worthwhile to train task-specific versions rather than perpetually calling external APIs.