A wave of decision models is entering the market that are designed for quick, low-cost classification, suited for cases when the slow, costly natural language generation of conventional LLMs is unnecessary. FLock.io has released THIS/THAT: a decision model built on Qwen3.5-2B and trained on Alibaba Cloud, small enough to fit on a laptop that can run without sending data to the cloud, making it suitable for regulated sectors where data privacy is non-negotiable.
LONDON – FLock.io, a privacy-focused AI company, has released an open-source “decision model” built on Alibaba Cloud’s Qwen3.5-2B that is small enough to run locally on laptops, smartphones and other edge devices.
The model, called THIS/THAT, has 1.88 billion parameters and is designed for high-volume tasks such as classification and quick decisions, rather than open-ended conversation. FLock.io says it produces no generated output tokens, and that local inference takes about 31.4 milliseconds, against seconds for standard large language models (LLMs). It also estimates the model is less than 0.04% the size of conventional LLMs. The company says THIS/THAT has processed more than 2.5 million decisions since going live.
TypeSafe’s Jev popularised the concept of decision models, alongside OpenAI’s Decisions API, Microsoft-Decision-1, and open-source options like Cloudflare’s Clef.
Aimed at regulated sectors
FLock.io says that because the model runs on the user’s own hardware, with no central GPU servers or cloud connection, sensitive data never leaves the machine. It pitches this at finance, healthcare, defence and government, where it says relying on centralised cloud LLMs creates security risks and regulatory hurdles. It contrasts this with hosted decision APIs, which require data to be sent to cloud servers.
“For many critical enterprise tasks, generating open-ended text is an unnecessary cost and performance bottleneck,” said Jiahao Sun, CEO of FLock.io. “With THIS / THAT Model, we are demonstrating that specialised 2B-parameter models running locally can outperform massive cloud LLMs in speed, accuracy, and cost-efficiency – all while keeping data strictly on-device.”
Benchmarks
In third-party evaluations, version 1.2 scored 94.1% accuracy on 68 local-geometry questions and an 87.8% success rate across 1,710 complex decision-making tasks on laptop-based benchmarks. FLock.io says this improved on version 1.0, tested just a day earlier, and that the gains came from optimising training data rather than enlarging the model.
The model was trained on Alibaba Cloud using synthetic data generated with Qwen Max. FLock.io says it draws on its GraphSynth and CircuitSynth research papers, which have been accepted at the ACL 2026 conference.
More about FLock.io
FLock.io is an AI research and infrastructure company focused on federated learning and distributed AI. Its platforms include AI Arena, FL Alliance, FOMO and an API Platform. FLock.io’s relationship with Alibaba Cloud’s Qwen began in April 2025. The company describes it as the first time a major centralised LLM has been deeply integrated with a decentralised training platform.
