How Decision Models’ Faster and Cheaper AI Decisions Are Powered by REAX (SN92)

Imagine a company receiving 10,000 customer support messages daily. Each message must be classified and assigned to the right department. While AI models like ChatGPT can handle this, processing thousands of simple decisions can be expensive and time-consuming.

Decision Models offers a faster, more affordable approach using specialized AI models that make decisions without lengthy explanations. The platform is connected to REAX (SN92), a subnet designed to decentralize the computing infrastructure that powers these models.

What ‘Decision Models’ Do?

Decision Models is a platform that gives developers access to open-weight AI models designed to make quick, structured decisions through a single API. These models, also called System-one models, answer specific questions by selecting from predefined options, providing probabilities, or assigning scores.

For example, a customer sends a company this message saying, “I was charged twice for my subscription this month. Please refund the extra payment.”

DecisionModels can analyze the message and determine which department should handle it, assigning a 94% probability to billing, 4% to technical support, and 2% to general enquiries. The company can then automatically direct the complaint to its billing department.

Unlike traditional chatbots, Decision Models return structured answers that applications can act on immediately, without generating lengthy explanations.

What Models Are Available?

Decision Models offers several AI models designed for different tasks, allowing developers to choose based on their needs.

1. S1-pro handles more complex text-based decisions and can answer up to three questions in a single request. It uses vLLM-SR d3, a 26-billion-parameter open-weight model.

2. s1-fast is designed for quick decisions involving short text, supporting up to 16 possible answers per question.

3. s1-vision processes images alongside questions, making it useful for tasks such as classifying products from photographs.

4. quyet-1-large is yet another model for text-based decisions.

The platform also provides an LLM auto-router that helps select an appropriate model based on the difficulty and importance of a request.

All these models can be accessed through a single API compatible with Jev, allowing developers to integrate them into existing applications without building separate connections for each model.

A Jevolution Is Coming to Bittensor

How Fast and Affordable Is ‘Decision Models?’

One of Decision Models’ main features is the REAX, a specialized inference engine designed to process AI decisions faster. Unlike traditional LLMs that generate responses word by word, the engine evaluates questions and scores possible answers without producing lengthy explanations.

REAX is estimated to have processed 14.2 decisions per second, compared to 5.4 on a reference server using the same GPU hardware. This represents approximately 2.6 times higher throughput, with some configurations achieving more than four times the performance.

Decision Models also offers competitive pricing, starting at $0.019 per million input tokens for s1-fast and $0.080 for s1-pro, with output tokens free. Customers can choose between a lower-cost Peer-to-Peer tier and an EU tier that processes requests within the European Union.

What Is REAX (SN92) and How Does It Work?

REAX (SN92) is a Bittensor subnet designed to decentralize the computing infrastructure behind Decision Models. While Decision Models provides the API that customers use, REAX allows independent computing providers to host and serve the AI models.

On REAX, miners provide computing power to run supported models and process requests, while validators check the accuracy and speed of their responses. For example, when a miner classifies a document, validators can compare its answer against the expected result to verify its accuracy.

Miners are rewarded based on how quickly and accurately they process requests, encouraging reliable performance across the network.

Who Is Behind Decision Models and REAX?

Decision Models and REAX (SN92) were developed by Florian Standhartinger, an AI consultant and creator of JevBench, an independent benchmark for evaluating the accuracy, speed, and cost of decision models. He previously worked with Chutes (SN64), where he gained experience with decentralized AI inference on Bittensor.

His JevBench project has attracted attention from major AI companies, including Microsoft, which referenced the benchmark when introducing its Microsoft-Decision-1 model.

The announcement of Decision Models also attracted significant attention, generating over 913K+ views on X and ranking third among major AI release announcements tracked over the same two-day period, ahead of Meta’s Muse Spark Open announcement.

The development of REAX also received financial backing from Const, whom Standhartinger publicly thanked for funding the subnet.

Why Does ‘Decision Models’ Matter for Bittensor?

Decision Models creates a commercial opportunity for REAX by targeting businesses that need to process large volumes of AI decisions quickly and affordably. For example, a customer service platform handling one million support tickets monthly could use Decision Models to classify complaints, identify urgent cases, and determine which require human attention.

REAX miners could provide the computing power needed to process these requests. As more businesses adopt Decision Models, demand for computing resources on REAX could grow, creating earning opportunities for miners.

➛ Contribute GPU Compute Here.