Gen 1 Slides Opus 5 Level Decks At A Fraction Of The Cost

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

Join us for our inaugural conference, Forge 2026

Blog

Gen 1 Slides Opus 5 Level Decks At A Fraction Of The Cost Gen-1 Slides: Opus 5-level decks at a fraction of the cost

PUBLISHED 9/10/2026

Table of Contents TL;DR Building specialized intelligence for agentic work Why this is a reinforcement learning problem The training partnership Encoding Genspark's standard A curriculum, not a single run Keeping the run stable Matching Opus 5 at a fraction of the cost Considerations for teams running long-horizon RL

Table of Contents

Talk to the Fireworks training team

Get in touch

TL;DR

• Genspark partnered with Fireworks Lab to post-train MiniMax M3 into Gen-1 Slides, a model that plans, writes, and checks its own slide decks end-to-end. • Gen-1 Slides matches Opus 5 on deck quality on Genspark’s evaluation, at about 1/17 of Opus 5's input-token list price. Per finished deck, Gen-1 Slides costs about 90% less than Opus 5. In production it is on par with Opus 5 on every metric, and it cut low-rated decks from 18% to 3.6% over its base. • Genspark's research team defined what a good deck is, encoded it as an evaluation standard, led the design of the algorithms to train that standard into the model, and validated it on live traffic. Fireworks Lab managed the full training to get the model there: the reward engineering, the 100-plus experiments, and the systems work that keeps a 100,000-token trajectory numerically stable.

Aggregate internal-grader score, 200 real tasks. Comparison is primarily against Opus 5, the frontier model served at scale on this task; Claude Fable 5 / 5.1 shown for reference only. Building specialized intelligence for agentic work

Slides are one of Genspark's highest-volume agentic workloads. Users generate them for quarterly reviews, client pitches, and board updates, and they need output that holds up in front of an audience, not a draft they still have to fix by hand. Off-the-shelf models, including the best proprietary ones, produce decks that look right at a glance but fail on inspection: text overflowing its box, blank renders, invented content. On a closed model, Genspark couldn't tune for the quality the workload needed, and at over a trillion tokens a month, couldn't control the cost either. Owning the model meant owning both. Why this is a reinforcement learning problem

A deck isn't a single output you can hand a model to imitate. Producing a good one means working in a live workspace across dozens of turns: planning the arc, writing the HTML, rendering it, reviewing the render, catching a layout defect, and fixing it, over hundreds of thousands of tokens in one session. A finished deck records none of that work, only its result. Examples can show the model what a good deck looks like, but not the sequence of judgments that produced it, so imitation alone can't teach the behavior. That is what made this a reinforcement learning (RL) problem, and a hard one on three counts: Long horizons. Episodes run dozens of turns. Small per-token errors compound into broken layouts, truncated files, and abandoned decks. Judgment, not just correctness. A deck can be syntactically perfect and visually incoherent. The reward has to capture design quality, not just task completion. Credit assignment across turns. A layout decision on an early turn can look fine and only produce a broken render dozens of turns later. The signal has to reach back from the failure to the decision that caused it.

The training partnership

Clearing all three at once is what stood between Genspark and a model they could own. That is exactly what Fireworks Lab is built for: it brings the training and inference infrastructure, compute, and research talent of a frontier lab to a customer's hardest training problems, so they can compete on quality, cost, and performance. Its researchers embedded with Genspark's team to post-train MiniMax M3, an open-weight multimodal model, into Gen-1 Slides. Genspark shaped the objectives and algorithms. They brought their real production environment directly into training and defined the standard behind it: the judgment of what makes a deck good, the design principles behind it, and the process for sharpening that standard. They also led the design of the algorithms to train that standard into the model. Fireworks Lab codeveloped the algorithm and managed the training process. Beyond the infrastructure, their researchers were deeply involved in optimizing the process for efficiency and for alignment with Genspark's goals. They drove more than 100 experiments and read trajectories to catch the model gaming the score. In one case, the model raised its visual-design score while task completion slipped, so the total went up but the deck wasn't better. Keeping a run whose episodes run past 100,000 tokens stable enough to converge is as much a systems problem as a research one. Encoding Genspark's standard

RL optimizes toward the evaluation, so what it measures is what the model becomes. Genspark's methodology scores each deck across the quality dimensions that matter and penalizes the defects that ruin one in practice: a blank render, text past its margins, invented content. The scores came from live production behavior. As training progressed, quality scores rose while penalties fell. Reward across training. Gen-1 Slides climbs from the MiniMax M3 base to Opus 5's level over the run. A curriculum, not a single run

Fireworks Lab structured the training as a sequence of stages, each raising the context length and fixing the failure the previous stage exposed. Learn what good design looks like. Fireworks Lab began with an SFT pass on curated decks, giving the model the visual vocabulary of good slides before any RL. Start RL at short context. Long contexts early on produce long, error-filled rollouts and a noisy gradient, so they started RL on short episodes to keep the signal clean, and lengthened the episodes as the policy steadied. The first solid checkpoint generated better slides but still laid them out unreliably. Extend the context window. Once the reward carried a clean signal, they widened the window to let longer, richer trajectories into training, and layouts became reliable.

As trajectories lengthened, the binding constraint stopped being the reward and became the numerics of a very long episode. Keeping the run...