Florian Explains Why Decision Models Could Be Much Bigger Than Chatbots

Decision Models and REAX (SN92) have quickly become one of the biggest talking points on Bittensor. The launch attracted close to a million views on X, miners rushed into the subnet, and people immediately started asking the obvious questions: what exactly is being built here, why does it need Bittensor, and is this really as important as the attention suggests?

The TAO Daily recently explained how Decision Models works and how REAX provides the decentralized infrastructure behind it. But Florian Standhartinger’s conversation with Mark Jeffrey made it clear that the ambition goes far beyond simply serving faster and cheaper AI models.

Decision Models Are Not Trying to Replace ChatGPT

Decision models are a new category of AI models built to make fast, structured decisions. Decision Models, with capital letters, is Florian’s platform that provides access to these models and is powered by REAX (SN92).

The category itself is meant for tasks that do not need a large language model. Imagine 10,000 customer messages arriving every minute. You do not need ChatGPT to write a thoughtful response to every message. Sometimes you simply need AI to decide whether each one belongs in billing, support or sales.

Florian believes these smaller decision models could eventually handle millions of these choices inside software. Rather than replacing ChatGPT, they can take over jobs that are currently handled by rigid if/else rules or by LLMs that are far more powerful and expensive than the task requires.

That is where Decision Models, the platform, comes in. It gives developers access to these models through an API, while REAX provides the infrastructure for serving and improving them.

Because the models are built for making decisions instead of generating long responses, Florian said they can be roughly 100 to 200 times faster and cheaper than large general-purpose models in some use cases. That makes AI practical for many small, frequent decisions where using a frontier LLM would simply be too slow or too expensive.

This Could Make AI Much More Useful Inside Software

The interesting part is how many of these decisions exist.

A decision model could determine which AI model should answer a request, route customer support tickets, moderate content, score documents or classify what a camera is seeing. Because the models can operate much faster, Florian also sees opportunities in areas such as robotics and other real-time systems where waiting several seconds for an LLM response is not practical.

One of the first models developed through the project is an LLM auto-router. It looks at a request and decides whether it really needs an expensive frontier model or whether a smaller, cheaper model can handle it. Florian said a major German IT consultancy is already evaluating the model in a pilot.

This is what the LLM Auto Router model outputs in order to decide which LLM to route to.

Play with examples here: https://t.co/2U4ZMWxu4i https://t.co/mqInO5kWhs pic.twitter.com/CYVGlz9f57

— Florian S (@airesearch12) October 17, 2026

That could become a very important use case. AI companies increasingly have access to many different models with very different costs and capabilities. If a tiny, fast model can decide where every request should go, businesses can reserve expensive models for the tasks that need them.

REAX Is Becoming More Than an Inference Subnet

At first glance, REAX (SN92) looks like a decentralized inference network where miners provide GPUs to serve the models available through Decision Models. That is part of it, but Florian is also using the subnet to improve the technology itself.

Miners are competing to make the REAX inference engine faster, while separate competitions are beginning to focus on building specialized models for individual decision-making tasks. Instead of one general model trying to handle everything, the subnet could eventually produce many small expert models, each optimized for a specific job.

The idea is then to potentially place a router in front of those models. A request comes in, the router understands what kind of decision needs to be made, and it sends the task to the specialist best suited for it. Florian said miners are already training use-case-specific models and early results have been encouraging.

So REAX is not simply trying to decentralize where Decision Models runs. It could become a factory for continuously producing faster, cheaper and more specialized decision intelligence.

Why Bittensor Fits This Particularly Well

Frontier LLM development is increasingly a game for companies that can afford enormous GPU clusters. An independent developer cannot realistically wake up tomorrow and compete with OpenAI or Google on training the next trillion-parameter model.

Decision models are different.

Florian pointed out that these models are small enough for independent AI engineers to train and improve without needing billions of dollars in infrastructure. That has already created a surprisingly competitive open-source ecosystem. Within days of Jev introducing the category, dozens of open models appeared, and some open-weight models are already outperforming the original closed Jev model on public benchmarks.

That is almost exactly the environment Bittensor is designed for. Instead of one company hiring a research team, REAX can open the problem to miners around the world and continuously reward whoever finds a better model or a faster way to run it.

Florian ultimately wants participation to become simple enough that someone with a strong gaming computer could contribute spare compute and earn subnet rewards. REAX already had roughly 100 unique miners shortly after launching.

The Most Important Part Is That There Is Already a Product

There is plenty of hype around Decision Models right now, but one thing makes REAX different from many new subnet launches: the commercial product already exists.

Decision Models launched with a working API, pricing and paying customers. Florian even said customers had found the service and started paying before he publicly announced it. Today, developers can already access open-weight decision models through the platform, while REAX is being used to improve the infrastructure underneath them.

That does not mean REAX has already won this market. Decision models are still extremely new, enterprise adoption takes time, and many of the performance claims around this young category will need to keep being tested as more people use the technology.

But the opportunity is easier to understand after hearing Florian explain it.

The goal is not to build another chatbot. It is to make AI cheap and fast enough to exist quietly inside software making millions of small decisions that previously were too expensive, too slow or too difficult to automate intelligently.

If that becomes a major part of how software is built, REAX gives Bittensor a chance to be directly involved in improving, serving and competing in one of AI’s newest model categories.

➛ Watch the full conversation: