Fireworks Lab 与 Genspark 后训练 MiniMax M3 推出 Gen-1 Slides

Join us for our inaugural conference, Forge 2026

PUBLISHED 9/10/2026

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.

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.

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:

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.

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.

Fireworks Lab structured the training as a sequence of stages, each raising the context length and fixing the failure the previous stage exposed.

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

An RL update is only correct if the two engines in the loop agree. The inference stack samples a rollout and the trainer scores it, and the update is unbiased only when both assign the same probability to the same token. They run on different code paths, so that agreement is never automatic. On a short episode the difference is negligible; on a trajectory past 100,000 tokens it compounds into a biased gradient.

Fireworks measures that agreement at every step, and the measurement caught the run's worst failure, one that raised no error and left nothing else looking wrong. The trainer and sampler were pulling apart on long trajectories, and Fireworks Lab traced it to the token stream.

Tokenization does not round-trip cleanly, so the sequence the trainer scored differed slightly from the one the model had sampled, and the importance ratio was computed against the wrong tokens. That biased the gradient on every step. Aligning the two token streams fixed it, and token- and batch-level filtering cleared the rare extreme divergences that remained.

A failure like this raises no error and biases every update. Catching it takes a training stack where inference and training are one aligned system, not a trainer bolted onto a separate inference engine.

Gen-1 Slides is live today as the default in Genspark AI Slides' Standard mode. On Genspark's evaluation, it matches Opus 5: it leads on visual design and on aggregate score, and ranks first in eight of the nine grader columns spanning three independent graders, and in the top two in all nine graders , at about 1/17 of Opus 5's input-token list price .

Measured per finished deck, Gen-1 Slides runs about one-tenth of the cost of Opus 5. For a team producing 1,000 decks per month, that means a reduction in spend of roughly 90% (from about $4,200 to $400). The budget that used to serve one user at frontier quality can now serve roughly ten.

In production, Gen-1 Slides is on par with Opus 5 on every metric, and cut low-rated decks from 18% to 3.6% against the base model. Because it was trained on full agent sessions rather than a benchmark, the planning, self-checking, and error recovery it learned are what show up in front of users.

Considerations for teams running long-horizon RL Designing the reward is a research problem. Keeping the run stable enough to optimize against it is a systems problem, and as trajectories grow long, that is where the hardest and least visible failures live. Three of them are worth planning for.

Gen-1 Slides is specialized intelligence Genspark owns: a model tuned to their domain that matches the frontier on one of their highest-volume workloads. What makes it durable is the standard behind it. Genspark's definition of a good deck lives in the reward and the eval, so it carries to every base model they train and produces the next when a stronger base arrives.

Slides are one of several capabilities Genspark has brought in-house with Fireworks . With the evaluation in hand, the quality ceiling for each capability is theirs to raise, instead of being dictated by someone else. If you want to push a workflow past what off-the-shelf models can give you, talk to Fireworks Lab .