{
  "id": 14477588,
  "title": "FLock.io Says Qwen-Based Decision Model Outperforms Larger LLMs on Selected Tasks",
  "url": "https://urgent.news/2026/10/14/flock-io-says-qwen-based-decision-model-outperforms-larger-llms-on",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-14T15:32:54.000Z",
  "source": {
    "name": "Korea IT Times",
    "slug": "korea-it-times",
    "url": "https://www.koreaittimes.com/news/articleView.html?idxno=158144"
  },
  "original_language": "en",
  "account": "London-based FLock.io unveiled on October 13 its THIS/ THAT Model, a compact open-source AI model engineered for swift and economical decision-making instead of expansive text generation. Rooted in Qwen3.5-2B with support from Qwen and Alibaba Cloud, THIS/ THAT Model leverages synthetic training data and is intended for tasks with predetermined outcomes, such as yes-or-no decisions and data classification. Since its launch, FLock.io reported processing 2.5 million decisions in three weeks.\n\nThe model targets enterprise applications where generating natural language incurs unnecessary processing time and cost, and where decisions can be directly returned. FLock.io cited System One or decision models, similar to TypeSafe's Jev, as notable in enterprise AI since September. Unlike conventional LLMs that produce open-ended prose, these systems choose from a set of predefined answers, embodying the psychological distinction between rapid, automatic \"System 1\" thinking and slower, deliberative \"System 2\" thinking.\n\nIn testing, version 1.2 of THIS/ THAT Model achieved an 87.8% success rate across 1,710 complex decision-making tasks, outperforming a benchmark of 40.6%. The release did not specify the baseline model configuration. For budget-strained optimization tasks, the model demonstrated a success rate four times that of GPT-5.6 and five times that of Claude Opus 5.\n\nFlank.io's CEO Jiahao Sun attributed the improvement to optimizing training data rather than increasing model size. The model, with 1.88 billion parameters, generates no output tokens and can run on standard developer hardware and edge devices, keeping data strictly local. Its Qwen3.5-2B foundation supports local deployment on devices like laptops and smartphones, with local inference latency of approximately 31.4 milliseconds, faster than seconds for standard LLMs.\n\nThe model's parameter count is estimated at less than 0.04% of Claude Opus 5, though FLock.io did not clarify this estimate. Training occurred on Alibaba Cloud using synthetic data from Qwen Max, informed by GraphSynth and CircuitSynth research papers. FLock.io highlighted the potential for specialized models to achieve high inference efficiency and strong performance on specific decision-making tasks, combining Qwen foundation models, Alibaba Cloud infrastructure, and FLock.io's research.",
  "summary": "London-based FLock.io announced on October 13 that it has introduced THIS / THAT Model, a small, open-source AI model designed for fast, low-cost decision-making rather than open-ended text generation. Built on Qwen3.5-2B with support from Qwen and Alibaba Cloud, the model uses synthetic training data and is intended for tasks with predefined answers, such as yes-or-no decisions and data…",
  "key_points": [
    "FLock.io launches THIS/ THAT Model on October 13",
    "Model outperforms larger LLMs like GPT-5.6 and Claude Opus 5",
    "THIS/ THAT Model processes 2.5 million decisions in three weeks"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}