If you follow artificial intelligence news, you have certainly heard of LLMs (Large Language Models), the technology behind tools like ChatGPT and Google Gemini. However, the AI universe has expanded dramatically. Today, engineers and researchers are developing highly specialized architectures to solve problems that pure text simply cannot address.
Below, we explore the main new acronyms defining the current landscape of artificial intelligence: VLM, SLM, LAM, and more.
1. VLM (Vision-Language Models)
- What they are: Models trained to simultaneously process and connect two primary modalities: images (or video) and text.
- How they work: Unlike a simple image classifier or an OCR (optical character recognition) tool, a VLM understands visual context and links it directly to natural language.
- Use cases: Analyzing medical images alongside medical reports, explaining complex infographics, describing videos frame-by-frame, and assisting visually impaired users.
- Examples: GPT-4o, Google Gemini 2.5 Pro, LLaVA, and Qwen-VL.
2. SLM (Small Language Models)
- What they are: Compact language models—typically between 1B and 14B parameters—designed to offer high performance using significantly fewer computational resources.
- Why they matter: While traditional LLMs require vast data center server clusters, SLMs can run locally on consumer devices like smartphones, laptops, and edge devices.
- Key benefits: Enhanced privacy (data remains on-device), offline availability, ultra-low latency, and drastically lower operational costs.
- Examples: Microsoft Phi-3.5, Llama 3.1 8B, Google Gemma 2, and MiniCPM.
3. LAM (Large Action Models)
- What they are: Models built not just to generate text or converse, but to take actions inside user interfaces (UIs) and computer applications.
- How they work: LAMs observe human interaction with software (button clicks, field inputs, page navigation) and translate natural language instructions into concrete steps within web or desktop interfaces.
- Use cases: End-to-end task automation, such as booking flights across complex web forms, filling out ERP systems, or operating legacy software without APIs.
- Examples: Rabbit OS (LAM), Adept’s ACT-1, and specialized web-agent architectures.
4. LMM (Large Multimodal Models)
- What they are: Generalist models trained natively to ingest and generate multiple data formats simultaneously—including text, audio, image, video, and code.
- LLM vs. LMM: Traditional LLMs process text natively and rely on external translation tools (like separate speech-to-text models) for other modalities. LMMs process multiple modalities within a single unified architecture.
- Why they matter: They enable fluid, real-time voice conversations, cross-modal reasoning (e.g., converting a hand-drawn wireframe directly into code), and native audio/video comprehension.
- Examples: GPT-4o, Google Gemini, and Claude 3.5 Sonnet.
5. LFM (Large Foundation Models) / World Models
- What they are: Advanced architectures that extend beyond language processing to build an internal representation or simulator of physical reality (“world understanding”).
- How they work: By observing video, spatial data, and physics interactions, these models predict future states of physical environments, understanding mechanics, spatial depth, and object persistence.
- Use cases: Autonomous driving, robotics, spatial computing, and complex scientific simulations.
- Examples: OpenAI Sora, Runway Gen-3, and physical AI platforms from NVIDIA and Wayve.
Summary Comparison
| Model Type | Primary Input/Focus | Key Advantage |
| LLM (Large Language Model) | Text / Code | Deep reasoning and broad text capabilities. |
| VLM (Vision-Language Model) | Image + Text | Visual comprehension aligned with natural language. |
| SLM (Small Language Model) | Text / Code | Runs locally on edge devices with low latency. |
| LAM (Large Action Model) | User Interface (UI) | Automates tasks across applications and web forms. |
| LMM (Large Multimodal Model) | Text, Audio, Video, Image | Unified cross-modal reasoning. |
| World / LFM | Spatial & Physical Data | Simulates physical reality and spatial dynamics. |