{
  "id": 16737325,
  "title": "No signal, no problem: Smaller models bring more AI capabilities on-device",
  "url": "https://urgent.news/2026/10/27/no-signal-no-problem-smaller-models-bring-more-ai-capabilities-on",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-27T10:34:46.000Z",
  "source": {
    "name": "Computerworld",
    "slug": "computerworld",
    "url": "https://www.computerworld.com/article/4239203/no-signal-no-problem-smaller-models-bring-more-ai-capabilities-on-device.html"
  },
  "original_language": "en",
  "account": "The release of OpenAI's GPT-3 in 2020 marked the introduction of the largest language model ever created, boasting 175 billion parameters. Since then, AI systems have continued to grow in size, with today's largest models surpassing a trillion parameters. However, a parallel trend is developing in AI research, focusing on creating smaller and more efficient models.\n\nPioneers like Tether are leading this compact AI movement, developing lightweight models that require significantly less computing power than their larger counterparts. These models are driving a shift toward on-device AI, enabling a wide range of applications to run locally without needing an internet connection.\n\nFor instance, Tether's suite of compact models ranges from four billion parameters to those small enough to run on a basic smartphone. Even the largest of these models is far smaller than the massive AI systems known as \"mega models.\" Research has shown that while larger models once offered clear advantages, the benefits of scaling have started to plateau, and diminishing returns are now becoming apparent.\n\nTether's approach to building smaller, more efficient models involves data curation and compression techniques. Their QVAC research team has focused on creating highly specialized, synthetic training datasets tailored to specific domains, such as medical reasoning, translation, and visual processing. These datasets contain 191 billion tokens across 19 STEM fields, ensuring the AI models can develop expertise in those areas.\n\nTo optimize model performance, Tether employs quantization, a technique that replaces high-precision weights with leaner, low-bit ones, reducing memory requirements. This technique has become the standard in the open-source AI community, with models as small as 4-bit models now delivering impressive results. The company's QVAC SDK further enhances the efficiency of on-device AI applications, offering tools like TurboQuant to compress model memory and TurboVec to optimize data retrieval.\n\nAs the AI industry grapples with the limits of scaling and diminishing returns from brute force, the rise of compact AI models presents a promising alternative. These smaller models can now deliver powerful capabilities on consumer devices, offering a more efficient and accessible approach to AI technology.",
  "summary": "When OpenAI’s GPT-3 was released in 2020, at 175 billion parameters, it was the largest language model the world had ever seen. In the years that followed, frontier AI systems exploded in size, with today’s largest models clocking in at well over a trillion parameters. But there is a countercurrent of AI research moving in the opposite direction, with pioneers like Tether building a new…",
  "key_points": [],
  "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."
}