{
  "id": 16776569,
  "title": "Build specialized multimodal models with LFM3",
  "url": "https://urgent.news/2026/10/27/build-specialized-multimodal-models-with-lfm3",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-27T17:16:52.000Z",
  "source": {
    "name": "Hugging Face",
    "slug": "hugging-face",
    "url": "https://huggingface.co/blog/LiquidAI/lfm3"
  },
  "original_language": "en",
  "account": "Liquid Foundation Models (LFM3) is the latest generation of models developed to serve as the foundation for specialized intelligence. These models integrate text, vision, audio, omni, and decision capabilities onto a shared backbone, all within a 2B parameter range. This allows users to select the appropriate model based on their modality, fine-tune it using their own data on a single GPU, and deploy it wherever their application is running.\n\nThere are four main changes compared to the previous LFM2 family. For LFM3, Liquid used their hardware-in-the-loop architecture search to redesign the architecture specifically for agentic use cases, which require more capability per parameter and fast inference over long contexts. Unlike its predecessor, LFM3-1.7B-Base comes in 14 device and backend combinations, ranging from phones and Raspberry Pi 5 to laptops, cloud CPUs, GPUs, and Qualcomm Hexagon NPU. With 24 layers, LFM3-1.7B-Base achieves comparable performance to LFM2-1.2B on short prompts across edge devices. However, when given 8K tokens in context, it is about 25% faster.\n\nEach of the LFM3 models branches from LFM3-1.7B-Base. Rather than running a single large general-purpose model, users can choose the one that fits their inputs, customize it, and run it on-device. The vision and audio models also receive two new in-house encoders, which are integrated into transformer-based architectures.\n\nLFM3-1.7B is ideal for local agents, tool calling, structured outputs, or domain-specific fine-tuning. It has been post-trained to run in agentic harnesses such as Hermes Agent, Pi, and OpenClaw, outperforming its predecessor, LFM2.5-2.6B, with 35% fewer parameters. LFM3-VL-2B is the first LFM vision model to take video as input, performing step-by-step reasoning before providing answers and identifying objects to understand their placement, offering strong potential for robotics. Despite having fewer parameters (1.9B against 3.1B), LFM3-VL-2B improves the average across 62 vision benchmarks by 13.8 points, bringing it to par with Gemma 4 12B, a model significantly larger than itself.",
  "summary": null,
  "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."
}