Liquid AI

Liquid AI is an AI lab that develops the Liquid Foundation Models (LFM), small models built to run on devices such as phones and laptops, and the LEAP platform to take them there.

The LFM2 models were published with open weights on Hugging Face, with deployment guides for ExecuTorch, llama.cpp and vLLM, according to the company's blog. The LFM2.5 family is the current line: the documentation cites text, reasoning, Japanese, vision and audio variants.

LEAP, Liquid AI's on-device AI platform, includes the Leap SDK, a Kotlin Multiplatform library that the documentation says runs on iOS, macOS, Android, desktop JVM, Linux and Windows, with a wasmJs target in preview.

What Liquid AI offers, then, is a small open model plus the deployment tools for teams that want inference on the device itself, instead of calling a remote API. For teams that prefer calling through an API, the company also mentions Liquid Console and a Playground to try models.

Liquid AI is for developers and product teams that need to run AI on the device itself, on phones, laptops or the edge, and for teams that want to fine-tune a small model with their own data.

The usage profiles supported by the official sources:

• Teams building mobile or desktop apps with local AI: the Leap SDK covers iOS, macOS, Android, desktop JVM, Linux and Windows, according to the documentation.
• Teams running models on laptops: the company announced native acceleration for AMD Ryzen and Ryzen AI processors in LEAP.
• Teams that need end-to-end audio: the LFM2-Audio model, according to its Hugging Face page, does not require separate speech recognition and speech synthesis components. The original version lists English as its language, and there is a Japanese variant.
• Teams that fine-tune models: Liquid AI recommends LFM2.5-1.2B-Base for tasks that require heavy fine-tuning and cites TRL and Unsloth for fine-tuning.

Prerequisites: the know-how to integrate a native SDK or a runtime such as llama.cpp, MLX or vLLM, and clarity about the license. The model cards on Hugging Face cite the LFM Open License v1.0, and it is the one that defines commercial use.

The main features of Liquid AI are the LFM family of open-weight models, the Leap SDK for on-device deployment and the options for fine-tuning and API access.

• LFM family: the company's blog describes text models with open weights on Hugging Face and the LFM2.5 line as the current one, with text, reasoning, Japanese, vision and audio variants.
• Small sizes: according to the blog, the first LFM2 models had 350 million, 700 million and 1.2 billion parameters, and the LFM2-VL vision model had 450 million and 1.6 billion variants, later one with 3 billion.
• Audio: LFM2-Audio-1.5B is an end-to-end speech model, and the LFM2.5-Audio version brings llama.cpp-compatible GGUF files for CPU inference.
• Leap SDK: a Kotlin Multiplatform library for iOS, macOS, Android, desktop JVM, Linux and Windows.
• Runtimes: the company publishes guides for ExecuTorch, llama.cpp, MLX and vLLM.
• Fine-tuning: TRL and Unsloth, plus the open liquid-finetune tool, which runs training on Modal.
• API access: Liquid Console and a Playground to try models.

The current list of models is on Hugging Face and in the Liquid AI documentation.

Liquid AI is worth it for teams that want to run AI on the device itself with small open-weight models. The limits lie in the license, in fast-changing versions and in the scope of small models.

Pros:

• Open weights on Hugging Face for the LFM2 models, with deployment guides for ExecuTorch, llama.cpp and vLLM.
• Leap SDK in Kotlin Multiplatform, covering iOS, macOS, Android, desktop JVM, Linux and Windows.
• Vision and audio models besides text, including an end-to-end speech model that dispenses with separate recognition and synthesis.
• Native acceleration for AMD Ryzen and Ryzen AI processors, announced for LEAP.
• Documented fine-tuning, with TRL, Unsloth and the liquid-finetune tool.

Cons:

• The license needs reading: the Hugging Face cards cite the LFM Open License v1.0, and older posts described different terms.
• Versions change: several LFM2 model cards point to replacements in the LFM2.5 line.
• The original LFM2-Audio version lists only English, and Japanese is in a separate variant.
• The fine-tuning path documented in liquid-finetune runs on Modal, a third-party GPU service.
• The wasmJs target of the Leap SDK, for the browser, is in preview, according to the documentation.

The main alternatives to Liquid AI are other labs that publish small or open models to run locally.

• Google, with the Gemma family: open-weight models from Google, with reduced-size variants.
• Microsoft, with the Phi family: small models published by Microsoft.
• Meta, with the Llama family: open-weight models in several sizes.
• Alibaba, with the Qwen family: open models in several sizes and modalities.
• Nous Research: Hermes open-weight models, plus Hermes Agent, which runs on your server.

What sets Liquid AI apart is its declared focus on on-device deployment, with the Leap SDK and LFM models that include vision and end-to-end audio. Teams that need large models to run on a server tend to look at the families above or at Nous Research, which publishes variants with hundreds of billions of parameters.