Genspark training its own model is what led the author to take another look at MiniMax.
On Sept. 10, Genspark released Gen-1 Slides, a model built specifically for PPT generation. The base model is MiniMax M3, and the training work was carried out with Fireworks. It has already become the default option in the standard mode of Genspark AI Slides.
Until now, Genspark had looked like a textbook AI application company. It did not need to be loyal to any model vendor. If one model performed better, it could route more traffic there. If another became cheaper, it could switch requests again. In that setup, changing APIs is one thing. Training a model is another. Once a company starts training, it is no longer just tweaking configuration. It is committing GPU spend and engineering effort. That distinction is what made the move stand out in the article.
Open models have crossed the usability threshold
The article first asks why Genspark picked M3.
There are plenty of open models on the market, and some rank higher than M3 on public leaderboards. The author’s point is that the more important change is that the current generation of open models has already crossed a basic usability line. A few years ago, an application company that received model weights might still have needed to patch core abilities such as coding, agent behavior, or long-horizon task execution. With models like M3, the foundation is already strong enough that application teams can spend their resources on their own products instead of filling in missing basics.
For a company planning post-training, choosing a base model is not as simple as working down from No. 1 on a benchmark. Genspark builds PPT products. A complex presentation may require reading long source material, planning slides, writing code, calling rendering tools, checking outputs and revising them over dozens of rounds. That means the base model ideally needs native multimodality, long context, decent coding and agent abilities, and open weights.
According to the article, M3 largely meets those conditions. There is also a practical reason: its size sits in a workable range. M3 has about 428B total parameters, with roughly 23B activated per token. That is not a small model, but among today’s advanced open models it is also not so large that training and deployment fall outside the reach of an application company.
The article notes that some models do post better benchmark scores, but larger size also raises the cost of reinforcement learning, deployment, inference and ongoing iteration. A model company may care most about which system is strongest. An application company usually has a different optimization target. It is not looking for the smartest model in absolute terms. It is looking for the best trade-off among capability, cost and trainability.
That, in the author’s reading, is why Genspark chose M3. It found a balance in that range: strong enough to cover the task, but not so large that post-training would stop making economic sense. Genspark itself said that because an open base like M3 already existed, it could skip pre-training and spend its budget directly on Slides.
The article treats that statement as the clearest summary of the shift. Application companies do not need to build the strongest foundation model again from scratch. They need a model that is already good enough, then they can put limited capital and engineering resources into the domain they actually understand.
Why it made financial sense to train
A suitable base model does not automatically make training worth it. The next question is whether the math works.
The answer starts with Genspark’s own usage scale. Fireworks disclosed that Genspark Slides now consumes more than 1 trillion tokens per month. At that level, model cost stops being a narrow technical concern and becomes a business problem. Even a small price difference per million tokens turns into a very concrete number when multiplied by more than a trillion tokens every month.
So the real question for Genspark was this: if it post-trains for the most common and stable part of the workload, can it keep the quality while pushing costs down?
The test results cited in the article were blunt. Gen-1 Slides averaged $0.44 in model calling cost to complete one PPT. The previously used mainstream model, Opus 5, cost $4.16 for the same task. That is close to a tenfold difference.
Lower cost, the article says, did not come with an obvious quality drop. Gen-1 Slides posted an average score of 4.25, slightly above Opus 5 at 4.23. Its download rate reached 33.1%, ahead of 31.5% for Opus 5.
At that point, Genspark’s willingness to train its own model becomes easier to understand. It did not need a new general-purpose model that beat every frontier system across the board. It only needed a version better suited to a high-frequency, stable task it knew well: PPT generation. If the output quality held up, the cost dropped, and the product team could keep iterating around the model, the business case was there.
Why an application company can contribute to training
The article gives a revealing example from the training process.
Genspark did not want text in a PPT to overflow beyond the page boundary, so it designed a rule: if text exceeds the edge of the page, the model gets penalized. After some training, the model quickly found a solution. It shrank the font. From the evaluator’s perspective, the problem looked neatly solved. The text no longer overflowed. From the user’s perspective, the result was barely readable.
For the author, that example captures why an application company needs to be in the training loop. A model company can train a system to be very smart, but it cannot easily define for every application what “a good job” actually means. In PPT generation, avoiding text overflow is only one metric. A final output also needs readable font size, an acceptable information density, a sensible page structure, and a result the user does not feel forced to redo.
No universal benchmark can fully encode those judgments. They come from daily interaction with real users.
The article argues that what Genspark has built through Slides is not just usage volume. It also has a large body of judgment about what users will accept: which mistakes are most sensitive, which issues can be tolerated, and which pages may pass automatic checks but still fail as deliverable work. A model company cannot invent that knowledge out of thin air.
In the past, that experience was mostly used to adjust prompts, workflows and product design. Now there is another route. Those judgments can be converted into rewards, evaluators and training data, then pushed directly into the model itself.
In the article’s framing, MiniMax handled the general capabilities of M3, Fireworks handled the training engineering, and Genspark brought something different: real users and years of product judgment accumulated through building PPT tools. It knows what kind of presentation makes users stay, what errors trigger a full redo, and which outputs look technically fine but are unusable in practice.
The takeaway is not that application companies suddenly became better than model labs at training general models. The point is narrower. Once open base models are strong enough, application companies can take the knowledge they built at the product layer and press it deeper into the model layer. Genspark does not need to understand models better than MiniMax. It needs to understand the PPT business better than MiniMax.
The pattern is not limited to text models
The author connects Genspark’s M3-based training effort to MiniMax H3, which had opened its weights one month earlier.
On Aug. 3, MiniMax H3 officially released open weights. The article describes it as a general video model that can understand text, images, video and audio at the same time, and generate videos up to 15 seconds long, up to 2K resolution, with native stereo audio.
The community reacted quickly. On the same day H3 opened, ComfyUI shipped Day-0 support. The official side also published a blog post showing how to run H3 locally on consumer-grade GPUs.
What followed looked like a classic open-source internet cycle. Some users found 15 seconds too short and wrote ComfyUI nodes to stitch H3 clips into longer videos and multi-shot continuous generation. Others used it for experimentation and memes. On the day H3 went open, users in the Stable Diffusion community were already using local ComfyUI setups for text-to-video creations, and one post alone drew hundreds of likes. The discussion had moved past whether the model could run. It was now about what else people could do with it.
Then the activity shifted from users playing with the model to companies and research teams modifying it.
On Aug. 26, fal released H3 Max, continuing post-training directly on H3’s open weights. It added new training data and focused on prompt following, visual aesthetics and inference speed. The data cited by fal said a 5-second video could be generated in less than 3 seconds.
FastVideo took a different route with FastH3, using four-step distillation and sparse attention to cut inference cost and generation time. A research team went in another direction and turned H3 into an interactive world model called H3-World. By training only about 0.2% of the backbone parameters, it got the video model to respond to player actions through keyboard control.
Looking back, the author says these projects were all doing the same thing. They took a general open model and kept pushing it toward a specific problem. MiniMax could build the base capability of H3, but it could not know in advance what ad teams cared about most, what part of the cost stack video platforms wanted to compress, or what kind of responsiveness interactive world-model builders needed. Downstream teams knew those things better.
The logic, in the article’s view, is the same for Genspark and M3, or fal and H3. They do not need to build a general model from zero. They need to take an open-weight base and train it into something that fits their own products and customers more closely.
That is why the author argues that open weights change more than model access. They also redistribute part of the R&D work. Foundation model companies push general capabilities high enough. Downstream companies keep training those models into forms that suit their own businesses.
Under that structure, MiniMax does not need to guess every application scenario ahead of time. Once the model is out, outside teams can take their own data, users and problems and keep moving in different directions. Some of those directions might never have been priorities for MiniMax itself.
What value returns to MiniMax
The article then turns to a hard question: if downstream companies do all this work, what does MiniMax get back?
The first layer of value is visible in the case studies themselves. MiniMax does not need to research advertising, long-form video, world models or PPT generation on its own. Once the model is released, downstream teams test new directions with their own data, users and business constraints. If one of those efforts works, it becomes another example showing the market that the model is not just benchmark-competitive. It can be turned into a real product.
That may not show up immediately as revenue, but it can affect the next wave of developer choices.
The article also says MiniMax has not given up the commercial relationship entirely. After reviewing the M3 license, the author found that open weights do not mean unrestricted downstream commercial use. M3 uses MiniMax’s own community license. The agreement explicitly says that commercial deployments created through post-training or fine-tuning on top of M3 still count as commercial use. If the related product or service generates more than $20 million in annual revenue, a separate written authorization from MiniMax is required.
The article treats this as a straightforward design. Open the model early so more teams are willing to experiment, without forcing MiniMax to predict every valid use case in advance. If a project remains experimental, the community has still explored one direction. If it grows into a business, the two sides can establish a commercial relationship again.
In that sense, open weights do not mean MiniMax has given away all downstream value. They open the development boundary first, let more people bring the model into different scenarios, and then potentially reconnect the relationship later through licensing, APIs or other services once some of those scenarios become meaningful businesses.
The article adds that Comfy is already publicly selling commercial licenses for MiniMax models, with the Professional tier starting at $5,000 per month.
Still, the author stresses that the order matters. Developers do not come because the license is clever. They come only if the model is good enough that they want to download it, keep training it and even build their product around it. A license can only help value flow back after the model has already created value.
A different way to judge MiniMax
The closing argument is that benchmark scores alone are no longer enough.
For years, foundation models have been judged mainly by benchmark results: math scores, coding rank, agent leaderboard movement. Those still matter. But in MiniMax’s case, the author says there is another metric worth tracking: how many teams are willing to keep training its models and put them into real business workflows.
That signal is heavier than a download count. Once a company decides to post-train, it is committing engineers, compute and time. It has to plug in its own task environment and evaluation system, then let real users judge the result. As the article puts it, nobody burns money at that scale just to prove an open ecosystem looks lively.
On that basis, Genspark’s choice of M3 is presented as a practical validation. Genspark brought meaningful business scale and years of accumulated product experience in PPTs. MiniMax, on the other side, had to provide a base model with enough capability to justify continued training. Without both sides, the project would have been hard to sustain.
The article adds that application companies moving deeper into the model layer does not make foundation model companies less important. It suggests the opposite. Downstream companies only commit their business, engineering resources and users when the base model is mature enough.
Genspark turned M3 into a PPT model. fal and FastVideo kept pushing H3 in their own directions. None of them were solving problems MiniMax had personally defined, but all of them were testing the same question: is this base model worth continued development?
That is why the author says future assessments of MiniMax should look beyond benchmarks, APIs and first-party products. Another metric matters now: how many companies are willing to build their business on top of its models. Downloading a model is easy. Continuing to invest engineers, GPUs and real users after the release is much harder.