The rise of decision models, and why they belong on edge computing
A wave of decision models has swept through the tech world in recent weeks. Typesafe’s Jev model popularised this new phase of enterprise AI in September, introducing a model that produces structured outputs cheaply, quickly and consistently. Others have followed such as OpenAI’s Decisions API, Microsoft-Decision-1, and open-source options like FLock.io’s THIS/THAT Model and Cloudflare’s Clef.
System One models, also known as decision models, are a lightweight, fast alternative to big, slow LLMs. Unlike conventional LLMs, they do not write text. The output can be a simple yes or no, or other classifications worked out through probability estimation. The name comes from the psychological framework of “system 1” thinking (fast, automatic) versus “system 2” thinking (slow, effortful).
There is a lot of buzz over the speed and low cost. But the real value is captured in being small enough to be deployed at the edge, on local or distributed hardware. Certain decision models are designed in a way that means no GPU server or cloud is necessary. Data never has to leave the device, whether it is a laptop or a smartphone. This allows them to be fully private and sovereign. For governments, banks and healthcare and other sectors carrying sensitive data, this is a game-changer.
The problem with conventional LLMs
Since the breakout year of generative AI in 2023, the goal had been to scale LLMs into enormous models that wrote paragraphs of natural language text. These chatbots excelled at complex reasoning, broad knowledge retrieval and unexpected tasks. But the bigger they grew, the slower they got (known as high latency), and the more computing power, memory and energy they demanded. The token overhead piled up, meaning costs rapidly rose. Moreover, the majority of frontier LLMs are centralised. This means data is sent to central servers for processing, introducing compliance risks.
What is a classifier/decision model?
There’s a reason decision models have attracted so much attention. AI agents are constantly making decisions where an elaborate explanation from an expensive model is unnecessary.
Sometimes all you need is a simple, reliable answer delivered virtually instantaneously. A decision model is a fraction of the size of a conventional LLM (less than 1% of the parameters). This is all that’s needed to make snappy classifications to help agents decide how to act, based on certain probabilities.
For example, you click “pay now” on a high-volume e-commerce website. The system must instantly evaluate whether to approve it, flag it for two-factor authentication, or deny it and flag potential fraud. A standard cloud LLM takes seconds to generate a natural language response in order to work it out, and burns huge API costs on token generation. In addition, the customer data is sent to an external centralised server, introducing data privacy risks.
A decision model, by contrast, is small enough to run locally, removing the need for the cloud. It can make an output in milliseconds, skipping natural language generation altogether with zero-token output.
Decision models belong at the edge
TypeSafe AI offers Jev primarily through a hosted API, which means data is sent to a central server for processing. In sectors like healthcare, finance and defence, this poses compliance risks.
Similar models like FLock.io’s open-source THIS/THAT Model (v1.2), built on Qwen3.5-2B, can run locally while still offering high performance. THIS/THAT is lightweight (just 1.88 billion parameters, which is less than an estimated 0.04% the size of Claude Opus 5). This means they can be run on standard developer hardware, smartphones and edge devices, keeping data local and latency ultra low. Edge computing allows decision models to run completely offline with zero dependency on external cloud providers, ensuring ultra-low sub-31ms latency, data sovereignty and operational resilience.
