YFarmX logoYFarmXIntelligence for an accelerating worldBriefings
YFarmX logoYFarmX
YFarmX

Liquid AI

LFM

Downloadable models for local applications

Released 4 August 20261 min readLarge Language ModelsLast updated:

Editorial illustration: LFM

Key facts

Hybridconvolutions and attention
Architecture
MIT spinoutLiquid AI
Origin
On-deviceedge hardware
Target
Efficiencyless memory, compute
Design goal

Liquid foundation models are AI models that developers can download and run on their own hardware. The current LFM2.5 family combines convolutions with attention.

Liquid models can run on your own hardware

Liquid foundation models are AI models that developers can download and run on their own hardware. The family includes models for text, pictures and audio. The current LFM2.5-24B-A2B reasoning model was released on 13 October 2026.

The current design combines two processing methods

Liquid’s technical release describes a hybrid of short convolutions and grouped-query attention. Different layers process information in different ways. The design aims to reduce the memory and computation needed for local use.

The right model depends on the task, available memory and supported runtime. Test the complete application, including any connected tools: local model inference alone does not establish where every tool sends data.

Downloadable weights have licence conditions

Liquid publishes weights on its official model collection. Its licensing page sets out the revenue threshold for paid commercial deployment and separate research, education and non-profit provisions. Check the checkpoint’s licence before using or redistributing it.

All LLMs →