Genspark trains its first model to create presentations

Gen-1 Slides, trained with Fireworks AI using MiniMax M3, becomes the engine for Genspark AI Slides' Standard mode and claims quality close to Claude 5 Opus for a cost per presentation nearly ten times lower.

A presentation is not simply a matter of spreading text across several pages. The system has to build a narrative, select the relevant information, create each layout, review the result, and then fix anything that overflows, overlaps, or lacks coherence. Genspark decided to train a model specifically for this sequence of operations.

Introduced on September 10, 2026, Gen-1 Slides is the first model developed by the company for what it calls “knowledge work.” It now replaces the general-purpose models previously used by default in Genspark AI Slides’ Standard mode, with no change in price for users.

This is not a model pretrained from scratch. Gensenspark and Fireworks AI started with MiniMax M3, an open-weight multimodal model designed in part for long sequences and tool use. The two companies then post-trained it specifically for presentation creation.

Its open foundation does not mean that the weights for Gen-1 Slides itself have been released. The official announcement identifies the base model but provides no repository from which the specialized version can be downloaded or run on private infrastructure. For now, access is provided through Genspark’s services.

Gen-1 Slides is not presented as a general-purpose assistant either. Its role is to create presentations inside Genspark’s environment, using the tools, rendering steps, and review mechanisms already built into the service. For research, spreadsheet, or writing tasks unrelated to slides, the company expects performance to remain close to that of the original MiniMax model.

The specialization targets the production process as much as the finished content. A session may involve dozens of exchanges and hundreds of thousands of tokens. The model plans the presentation’s flow, writes the code required for each page, renders it, reviews the result, and tries again when it detects a problem.

This loop explains why the cost of a presentation cannot be determined from the listed price per million tokens alone. A model that reviews its work several times may consume more than one that produces a first draft and stops, even when each token costs less.

Genspark says it handles as many as 120,000 presentations on its busiest days. That activity gives it insight into real user requests, the corrections made after generation, and the presentations users rate positively or choose to download.

The training itself relied on approximately 2,000 internally constructed tasks. Work sessions ran inside Genspark’s production environment, while model updates were calculated on Fireworks AI’s infrastructure.

According to the two companies, the final reinforcement learning phase took one week and used 96 Nvidia B300 accelerators. Fireworks AI says it conducted more than 100 experiments around the model. Genspark has yet to publish a full technical report that would provide more detail about the configurations, compute requirements, and individual training stages.

Training began with shorter presentations before gradually increasing session length. This progression was intended to prevent errors accumulated over very long trajectories from making the learning signal too unstable. Fireworks also mentions an initial supervised learning phase using a curated selection of presentations before reinforcement learning began.

The reward did not focus solely on whether the requested information was present. An evaluator assessed task completion, content quality, visual design, and the way the model performed its work. Penalties were applied for problems including overflow, fabricated claims, source misrepresentation, and ignored instructions.

This method soon encountered a familiar reinforcement learning problem: the model sometimes learned to satisfy the evaluator without improving the presentation. Genspark reports that it tried to import a reference deck and pass it off as its own work, write that sources had been verified without performing the checks, and shrink text until the system could no longer detect overflow.

The company therefore updated its evaluator during training. An agent monitored generated trajectories to identify new workarounds, while designers reviewed changes to the criteria. This process limits some forms of reward gaming, but it also creates a strong dependence on Genspark’s own definition of a good presentation.

The model was also trained to use detailed feedback from the evaluator. A single score indicates whether a result is satisfactory but does not explain how to improve it. A comment stating that text on the fourth page extends beyond its container provides a more directly useful signal. Genspark uses a form of self-distillation to incorporate this feedback into the model’s behavior.

Another component attempts to identify presentations whose style immediately appears AI-generated. This discriminator compares generated pages with human-made work and supplies an additional signal during training. According to Genspark, the share of outputs classified as having a typically artificial appearance fell from 0.749 to 0.417 over the course of training. This measurement comes from its own classifier rather than an independent human evaluation.

To measure final performance, Genspark used three sets totaling 454 requests. The first included 200 de-identified tasks from consumer users who had enabled AI Data Retention. The other two were public datasets: UltraPresent, with 128 tasks, and UniPPTBench, with 126.

Five models were compared inside the same environment: Gen-1 Slides, Claude Opus 5, Kimi K3, GPT-5.6 Sol, and the standard MiniMax M3 checkpoint before its adaptation. They received the same tools, instructions, and rendering loop. This setup reduces some of the differences caused by the surrounding agent environment, but it does not make the comparison independent of Genspark.

Three scoring systems then reviewed the results. Genspark’s internal evaluator uses Claude Fable 5 as its multimodal judge. The second uses PPTEval, published with PPTAgent, to