Models built for real-world constraints.

www.liquid.ai · 2026-08-04 · by Long Nguyen ·generic
The world’s most efficient foundation models — designed for on-device, edge, and cloud deployment.

Models built for real-world constraints.

Liquid Foundation Models

Models built for real-world constraints.

We’ve redefined what’s possible, designed for efficiency, speed, and real-world deployment on any device.

Build with LEAPView docs

From wearables to robotics, phones, laptops, cars, and more, LFMs run seamlessly on GPUs, CPUs, or NPUs, making intelligence accessible everywhere.

Explore LFMs

Frontier-grade intelligence at a fraction of the compute, on any hardware, on or off the cloud.

Every LFM is free to download, run, and fine-tune — including commercially — until your company passes $10M in annual revenue. See pricing & licensing

Customize your model with us

All familiesLFM2LFM2.5

Text Models

LFMs deliver powerful performance in a lightweight, customizable, and compute-efficient footprint for deployment in any environment.

  • LFM2.5-2.6B
    • DemoDocsBlog[](https://huggingface.co/blog/LiquidAI/lfm2-5-2-6b)
  • LFM2.5-1.2B-Base
    • DocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Base)
  • LFM2.5-1.2B-Thinking
    • DemoLEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking)
  • LFM2.5-230M
    • DocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-230M)
  • LFM2.5-350M
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-350M)
  • LFM2.5-8B-A1B
    • DemoDocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-8B-A1B)
  • LFM2.5-1.2B-JP-202606
    • LEAPDocs[](https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP-202606)
  • LFM2-2.6B
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2-2.6B)
  • LFM2-700M
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2-700M)
  • LFM2-24B-A2B
    • DocsBlog[](https://huggingface.co/LiquidAI/LFM2-24B-A2B)
  • LFM2-1.2BDeprecated
    • Blog
  • LFM2-350MDeprecated
    • Blog
  • LFM2-8B-A1BDeprecated
    • Blog

    Vision-Language Models

    Multimodal models using vision and text inputs and outputs with capabilities designed for low latency and device aware deployment.

  • LFM2.5-VL-1.6B
    • DemoLEAPDocs[](https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B)
  • LFM2.5-VL-450M
    • DemoLEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-VL-450M)
  • LFM2-VL-3B
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2-VL-3B)
  • LFM2-VL-450MDeprecated
    • Blog

    Audio Model

    End-to-end foundation model for audio and text generation. Designed for low latency, it enables responsive, high-quality conversations with only 1.5 billion parameters.

  • LFM2.5-Audio-1.5B
    • DemoLEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B)
  • LFM2.5-Audio-1.5B-JP
    • Docs[](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B-JP)

    Nano Models

    Tiny customized models for specific tasks and knowledge.

  • LFM2.5-Encoder-350M
    • DemoDocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M)
  • LFM2.5-Encoder-230M
    • DocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-Encoder-230M)
  • LFM2.5-1.2B-Instruct
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct)
  • LFM2.5-Embedding-350M
    • DocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-Embedding-350M)
  • LFM2.5-ColBERT-350M
    • DocsBlog[](https://huggingface.co/LiquidAI/LFM2.5-ColBERT-350M)
  • LFM2.5-VL-450M-Extract
    • Docs[](https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract)
  • LFM2.5-VL-1.6B-Extract
    • Docs[](https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B-Extract)
  • LFM2-2.6B-Exp
    • LEAP[](https://huggingface.co/LiquidAI/LFM2-2.6B-Exp)
  • LFM2-350M-Math
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2-350M-Math)
  • LFM2-350M-PII-Extract-JP
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2-350M-PII-Extract-JP)
  • LFM2-2.6B-Transcript
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2-2.6B-Transcript)
  • LFM2-350M-ENJP-MT
    • LEAPDocsBlog[](https://huggingface.co/LiquidAI/LFM2-350M-ENJP-MT)
  • LFM2-350M-ExtractDeprecated
    • Blog
  • LFM2-1.2B-ExtractDeprecated
    • Blog
  • LFM2-1.2B-RAGDeprecated
    • Blog
  • LFM2-ColBERT-350MDeprecated
    • Blog
  • LFM2-1.2B-ToolDeprecated
    • Blog

  • ShieldFlow (LFM PII)
    • Request accessTalk to our team
  • LFM2-2.6B-MMAI
    • BlogRequest accessTalk to our team

    Custom Models

    Purpose-built LFMs shaped around your data, use case, and constraints.

  • Don’t see the model you need?Request a model
  • Want to train an LFM on your own data?Explore fine-tuning

Built for efficient inference everywhere you need it.

Unmatched speed, quality, and memory-efficiency on the edge, or in the cloud.

Whether deploying on smartphones, laptops, vehicles, or any other device, LFMs run efficiently on CPU, GPU, and NPU hardware. Designed for millisecond latency, on-device resilience, and data privacy, LFMs unlock the full potential of local, cloud, and hybrid AI across industries.

Galaxy S26 UltraiPhone 17 ProMacBook ProGMKtec EVO-X2

Decode throughput

\~350M\~800M\~1.2B

Decode throughput on Galaxy S26 Ultra · SM8850 · CPU · llama.cpp b9050 · Android · Q4\_K\_M

Each device runs its platform's native runtime config.

Latest model releases

Read more

AUG 4, 2026ModelsLFM2.5-2.6B: Deploy Agents Everywhere

07.2807.28ModelsLFM2.5-Encoders: Fast at Long Context, Even on CPUJUL 28, 2026ModelsLFM2.5-Encoders: Fast at Long Context, Even on CPU

06.2506.25ModelsLFM2.5-230M: Built to Run AnywhereJUN 25, 2026ModelsLFM2.5-230M: Built to Run Anywhere

06.1806.18ModelsLFM2.5 Retrievers: Bi-directional LFMs for Fast Multilingual SearchJUN 18, 2026ModelsLFM2.5 Retrievers: Bi-directional LFMs for Fast Multilingual Search

www.liquid.ai · 2 images · 138 links

Comments (0)

No comments yet