{
  "id": 19484536,
  "title": "My 8GB laptop now runs an AI that can see and talk, and it never goes online",
  "url": "https://urgent.news/2026/11/11/my-8gb-laptop-now-runs-an-ai-that-can-see-and-talk-and-it-never-goes",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-11T11:00:22.000Z",
  "source": {
    "name": "XDA Developers",
    "slug": "xda-developers",
    "url": "https://www.xda-developers.com/my-8gb-laptop-now-runs-an-ai-that-can-see-and-talk-and-it-never-goes-online/"
  },
  "original_language": "en",
  "account": "My 8GB MacBook Air, a device usually insufficient for running sophisticated local large language models, has now been equipped with an AI model named LFM3-Omni. This model, boasting just 2.2 billion parameters, released on October 27, can process images, video, and speech and respond either in text or speech. What sets this model apart is its ability to perform all these tasks on the 8GB system without needing to connect to the internet.\n\nDeveloped by Liquid AI, LFM3-Omni is an open-weight model released under the LFM Open License v1.0. I initially attempted to run it via Ollama, but it only supported text and vision processing. To enable speech capabilities, I utilized a free community app called lfm3-mega-demo, built by a developer known as overment on Github. This app employs Liquid's official 8-bit build through Apple's MLX framework to display a web interface with tabs for talking, looking, transcription, and speech generation.\n\nWhile I was wary about potential sampling settings in the app, my checks revealed that the app adheres to the same greedy text decoding settings as the model itself. When running the app, I found that the peak memory usage for the app was around 2,968MB, with about 2,814MB dedicated to the model within the Mac's shared memory. This demonstrates the impressive efficiency of LFM3-Omni, leaving ample memory for other tasks.\n\nTo evaluate its vision capabilities, I tested the model by presenting various objects to the webcam through the Look tab. The model accurately identified a DualSense controller as a black PlayStation controller and a TKL keyboard as a mechanical keyboard, even when the latter was a compact board without a numeric keypad. During a numerical analysis test, I provided the model with a table of PC components and asked a series of questions. The model successfully extracted the correct information, with the exception of arithmetic errors, which were attributed to a design choice rather than a limitation of the model's vision capabilities.",
  "summary": "Small LLMs are doing less for more, and I'm here for it",
  "key_points": [
    "8GB MacBook Air runs LFM3-Omni AI model",
    "Model processes images, video, speech locally",
    "No internet connection required for operation"
  ],
  "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."
}