THIS / THAT Model is for high-speed, low-cost decision making. It’s a small, open-source decision-model that decides in 30 milliseconds for a millionth of a cent. It’s trained by FLock.io and is free to use, without vendor lock-in.

There is certainly a time and a place for LLMs. But THIS / THAT Model is optimised specifically for queries that have a predefined set of possible answers. In such cases, using an LLM is overkill. THIS / THAT Model is small enough to run on standard developer hardware of edge infrastructure, meaning your data can stay local and latency can stay ultra low.

We are excited to announce that the academic paper behind THIS / THAT 1.0 is now out, outlining the architecture and performance of this lightweight model. It’s authored by our Chief AI Scientist Zehua Cheng, Wei Dai and our founder and CEO Jiahao Sun in collaboration with the University of Oxford. This blog explains how it works and what our results show.

Model 1.0 is found here, however the newer version 1.2 has already been released (get it here: https://huggingface.co/flock-io/this-that-model-1.2)

Explore the code: https://github.com/FLock-io/this-that-model

[ 👋 Hi there! If you’re here to find out more about FLock.io, follow us on X and email us at hello@flock.io to learn what we do in distributed inference, federated learning and sovereign AI. See our other research and whitepaper.]

The problem with general-purpose LLMs

General-purpose LLMs are great at generative reasoning and complex logic. But sometimes that goes beyond what’s needed for a simpler task. Using LLMs for structured, binary or multi-choice classification presents major bottlenecks:

High latency

Generating full text responses takes time – far too long when all you really need for a given task is a structured label.

Token overhead

The extra or hidden token consumption that occurs in LLMs and AI agents before or during a task, independent of the actual user prompt or question. This means costs rapidly rise.

Checking the output

Does the model output need to be validated and cleaned to make sure it matches expectations?

An agent’s workflow is full of decisions. Which queue should a support ticket enter? Does an LLM output follow instructions and comply with safety and formatting instructions? Does a refund request meet the policy criteria? Sending each question to a general-purpose LLM adds generation time, token costs and a response the application must interpret.

To solve this, FLock.io has introduced THIS / THAT Model.

How THIS / THAT Model works

THIS / THAT is designed for questions with a defined set of possible answers, where those overheads add up across repeated calls. Instead of generating text auto-regressively token by token, it works through structured probability estimation.

You define the possible answers upfront. That’s where its name comes from: you declare the allowable answer set (THIS , THAT, or any list of candidate choices) prior to evaluation. The model evaluates the input just once, then returns probabilities over exactly those options. Multiple questions about the same input can share that pass, reducing computational overhead. You can also set confidence thresholds to send uncertain cases for review by a human or an LLM.

Unlike LLMs, the model generates 0 output tokens. It’s built on a 1.88B / 2B parameter model using a hybrid linear-attention (delta-rule) backbone, making 32k-token inputs inexpensive to process ($O(L)$ computational complexity). It has a schema-first prompt layout, allowing the prompt prefix (the questions and option schemas) to be computed once and reused across millions of inputs, cutting prefill costs significantly for high-volume request classification.

What our research shows

Our research paper shows how targeted architecture and specialised fine-tuning allow a lightweight model to dramatically outperform traditional baseline models on targeted evaluation sets:

• 94.1% accuracy versus Jev’s 76.5% on 68 local-geometry questions recorded by a third party.
• An improvement from 40.9% to 83.9% on 7,305 spatial questions after targeted training, with evaluation examples held out but task types represented in training.
• Sustains 32 decisions per second on a single 16 GB consumer/laptop GPU (consuming ~80 W), costing $0.000217 in electricity for a full 42-family benchmark pass (compared to 155 minutes and $10.63 for frontier LLM API calls).

It’s just 1.9B parameters, runs locally, and has a measured decision latency of ~31ms on an RTX 5080 Laptop GPU. FLock.io is the pioneering decentralised model training platform to give developers and AI agents decision-making capabilities they can run on their own hardware with THIS / THAT Model.

THIS / THAT Model 1.2 is open source, free to use and runs on your hardware, so your data stays home.

Get the model: https://huggingface.co/flock-io/this-that-model-1.2

Explore the code: https://github.com/FLock-io/this-that-model

More about FLock.io

FLock.io is an AI research and infrastructure company pioneering enterprise-grade federated learning and distributed AI solutions. Its decentralised federated learning architecture and production-ready platforms (AI Arena, FL Alliance, and FLock API Platform) enable organisations to train and deploy their own custom AI models on local hardware while maintaining full data privacy, model ownership, and regulatory alignment by design.

FLock.io is internationally recognised for its academic research, including the NeurIPS award-winning paper “FLock: Defending Malicious Behaviors in Federated Learning with Blockchain” and sponsors computer science PhD students at the University of Oxford.

Follow FLock on LinkedIn and X and email the team at hello@flock.io.