Built on Qwen3.5-2B, FLock.io and Oxford University researchers have launched an open-source AI decision model that outperforms Jev and larger LLMs in various benchmarking tests and can be deployed on local hardware to protect data privacy.

London, England, United Kingdom, October 16, 2026 -- FLock.io, a developer of privacy-preserving AI infrastructure, has launched THIS/THAT Model: an open-source, lightweight “System One” decision model built on the Qwen3.5-2B foundation model. It can be run locally, ensuring data sovereignty. It was developed with researchers from the University of Oxford and with the support of Alibaba Cloud.

The launch joins a wave of decision models that have been released in recent weeks, which provide a low-cost, small, high-speed alternative to chatbots where generating text is unnecessary. Designed for enterprise tasks that require structured choices rather than open-ended text generation, it has already processed 2.5 million decisions three weeks after going live. TypeSafe’s Jev popularised the concept, alongside OpenAI’s Decisions API, Microsoft-Decision-1, and open-source options like Cloudflare’s Clef.

THIS/THAT can be deployed locally for full data privacy and sovereignty

Setting THIS/THAT apart is how it is not only open-source, but also small enough to run on your own device, whether it is a laptop, smartphone, hospital gateway or rugged field tablet in a warzone. The 1.9-billion-parameter model requires no central GPU servers and no cloud connectivity. By executing decisions locally, it ensures sensitive data never leaves the user’s machine.

For regulated sectors such as governments, banking institutions, and healthcare providers, relying on centralised cloud LLMs introduces significant data security risks and regulatory hurdles. Unlike hosted alternatives, THIS / THAT Model guarantees total data sovereignty: no enterprise data is ever transmitted to OpenAI, Microsoft, Cloudflare, or third-party servers. 

Jiahao Sun, CEO of FLock.io, said: “For many critical enterprise tasks, generating open-ended text is a performance bottleneck and an unnecessary cost. With THIS / THAT Model, we have demonstrated that specialised 2B-parameter models running on local devices can outperform huge cloud LLMs in accuracy, cost-efficiency and speed– all while keeping data on-device.”

The benchmarking results

In third-party evaluations, version 1.2 achieved 94.1% accuracy versus TypeSafe Jev’s 76.5% on 68 local-geometry questions; 87.8% success rate on laptop-based benchmarks, outperforming Claude Opus 5’s 83.4%. In budget-constrained optimisation, its success was 4x that of GPT-5.6 and 5x Claude Opus 5. This was a rapid improvement from 1.0’s tests just a day before, achieved through optimising training data rather than increasing model size. It drastically reduces compute and latency bottlenecks compared to cloud-hosted LLMs.

See the latest benchmarking on Hugging Face: https://huggingface.co/flock-io/this-that-model-1.2

Source code repository: https://github.com/FLock-io/this-that-model  

Decision models address the LLM compute and latency bottleneck

Enterprise AI is witnessing the emergence of “System One” decision models, a specialised class of models. While conventional LLMs auto-regressively generate open-ended text, decision models process queries over predefined answer spaces (such as binary or categorical choices). Borrowed from Daniel Kahneman’s psychological framework, “System 1” represents fast, automatic pattern execution, contrasting with the slow, effortful “System 2” reasoning performed by generative chatbots like ChatGPT.

It addresses a core problem faced by AI agents’ workflow, which involves thousands of routine micro-decisions, such as approving transactions. Relying on general-purpose cloud LLMs for these deterministic tasks creates severe operational bottlenecks. They have high latency, with generating text token-by-token taking seconds per query. Token overhead is high, with hidden prompt-processing costs accumulating rapidly over millions of API calls. Apps must also continually clean and validate the text.

THIS / THAT Model’s technical specifications

THIS/THAT was built on Qwen3.5-2B: a tiny, edge-friendly foundation model that can be run locally on laptops and smartphones. It is lightweight, with only 1.88B parameters (less than an estimated 0.04% the size of Claude Opus 5) and zero generated output tokens. The local inference latency is ultra low, at approximately 31.4 milliseconds, compared to standard LLMs taking seconds. It was trained on Alibaba Cloud using synthetic training data generated using Qwen Max. It drew on key findings from FLock.io’s recent research papers, GraphSynth and CircuitSynth, which were accepted into the prestigious ACL 2026 conference. 

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, FOMO and the 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.

Qwen, a tier 1 LLM developed by Alibaba Cloud, partnered with FLock.io in April 2025. It marked the first time a major Web2 centralised LLM has chosen to deeply integrate with a decentralised training platform.

Website: https://www.flock.io/ 

Media contact: 

Phoebe Hennell 

London

[email protected]

Contact Info:
Name: Sheharyar Khan
Email: Send Email
Organization: HackerNoon
Website: https://hackernoon.com/

Release ID: 89206434

In the event of encountering any errors, concerns, or inconsistencies within the content shared in this press release, we kindly request that you immediately contact us at [email protected] (it is important to note that this email is the authorized channel for such matters, sending multiple emails to multiple addresses does not necessarily help expedite your request). Our dedicated team will be readily accessible to address your feedback within 8 hours and take appropriate measures to rectify any identified issues or facilitate press release takedowns. Ensuring accuracy and reliability are central to our commitment.