On September 11, AI application developer Genspark introduced Gen-1 Slides, a specialized model crafted for PPT creation, leveraging the open-weight foundation of MiniMax M3 and undergoing collaborative post-training with Fireworks AI. This model now serves as the default engine for Genspark's AI Slides standard mode.

Genspark noted that generating presentations demands the model to perform content structuring, page assembly, rendering verification, and iterative refinements, with the platform producing up to 120,000 decks daily. This extensive real-world workload and user input provided essential scenario expertise for the targeted PPT-focused training.

Building on M3's core capabilities, Genspark integrated accumulated task insights, setting final delivery quality as the reinforcement learning objective, with emphasis on enhancing layout precision, data credibility, and visual design. According to company disclosures, across nine evaluation setups spanning three dataset groups and three scoring systems, Gen-1 Slides secured first place in eight categories, with overall PPT quality aligning closely with that of Claude Opus 5.

On the cost front, in a test of 200 authentic tasks within identical runtime conditions, Gen-1 Slides incurred an average model invocation expense of roughly $0.44 per deck, compared to about $4.16 for Claude Opus 5, translating to roughly one-tenth of the latter's cost. M3's native multimodal capabilities, million-token long-context handling, and tool-calling stability addressed the central requirements of Genspark users for slide generation.

M3 employs a Mixture-of-Experts architecture with roughly 428 billion total parameters, lowering computational and deployment loads for post-training, thereby boosting iteration efficiency. In its Gen-1 Slides announcement, Genspark expressed gratitude toward the Chinese team at MINIMAX-W, acknowledging that the open-source nature of MiniMax M3 enabled the team to bypass pre-training, channeling resources and budget directly into PPT-specific training, which allowed model development to conclude within weeks.

Genspark intends to extend this methodology to spreadsheets, documents, and research tasks, exploring the fusion of open-source models with application-specific expertise to elevate delivery quality and cost-effectiveness across office-oriented workflows.