LFM (Liquid Foundation Model)
By Liquid
Liquid Foundation Models (LFM) are a new class of generative AI models developed by Liquid AI, designed for enterprises and developers needing efficient, scalable AI solutions.
Publisher review
Liquid Foundation Models (LFM) are a new class of generative AI models developed by Liquid AI, designed for enterprises and developers needing efficient, scalable AI solutions. These models are built from first principles using computational units rooted in dynamical systems, signal processing, and numerical linear algebra, making them adaptable for sequential data like text, audio, and video. The first generation includes 1B, 3B, and 40B parameter models, each optimized for specific use cases, from edge devices to data centers.
LFM2.5-1.2B-Thinking, the latest addition, is a 1.2B parameter model tailored for on-device reasoning, running entirely on phones with just 900 MB of memory. It outperforms larger models like Qwen3-1.7B in reasoning benchmarks, making it ideal for mobile and embedded applications. Liquid AI targets industries like financial services, biotech, and consumer electronics, offering private, edge, and on-premise deployments.
The models are optimized for hardware from NVIDIA, AMD, Qualcomm, Cerebras, and Apple, ensuring broad compatibility. LFM-1B and LFM-3B set new benchmarks in their size categories, with LFM-3B delivering performance comparable to 7B-13B models while being significantly smaller. LFM2.5-1.2B-Thinking introduces systematic reasoning capabilities, generating thinking traces before answers, which enhances its performance on agentic and reasoning-heavy tasks.
The model is available via Hugging Face, LEAP, and Liquid AI's Playground, with partnerships extending to Qualcomm, Ollama, FastFlowLM, and Cactus Compute for diverse deployment scenarios. However, the enterprise pricing model requires custom quotes, and while LFM excels in efficiency, not all benchmarks show uniform superiority over competitors like Meta's Llama or Google's Gemma. The trade-offs include limited transparency in pricing and varying performance gains across tasks, but for edge and on-device AI, LFM represents a significant leap in quality and efficiency.
How it works
-
On-device reasoning
LFM2.5-1.2B-Thinking runs entirely on phones with 900 MB memory, enabling offline AI tasks.
-
Hardware optimization
Optimized for NVIDIA, AMD, Qualcomm, Cerebras, and Apple hardware, ensuring efficient inference.
-
State-of-the-art performance
LFM-1B and LFM-3B achieve top benchmarks in their size categories, outperforming larger models.
-
Systematic reasoning
LFM2.5-1.2B-Thinking generates thinking traces before answers, enhancing reasoning-heavy tasks.
-
Scalable deployment
Supports private, edge, and on-premise deployments across industries like biotech and finance.
-
Efficient memory footprint
LFM-3B delivers 7B-13B model performance with a smaller memory footprint, ideal for edge devices.
-
General-purpose AI
Models sequential data like text, audio, and video, adaptable to diverse applications.
Strengths and trade-offs
Strengths
- LFM2.5-1.2B-Thinking outperforms Qwen3-1.7B in reasoning benchmarks while using less memory.
- LFM-3B matches Phi-3.5-mini performance despite being 18.4% smaller, ideal for mobile applications.
- LFM-1B achieves the highest MMLU (5-shot) score (58.55) among 1B models, surpassing transformer-based alternatives.
- LFM2.5-1.2B-Thinking runs entirely on-device with 900 MB memory, enabling offline reasoning tasks.
Trade-offs
- Enterprise pricing requires custom quotes, lacking transparent tiered plans for smaller teams.
- Not all benchmarks show consistent performance gains over competitors like Meta's Llama 3.2.
- Limited public documentation on model architecture and training datasets for independent verification.
- Deployment flexibility comes with complexity, requiring hardware-specific optimization for peak performance.
Pricing context
Custom quote for enterprise solutions
Getting started with LFM (Liquid Foundation Model)
-
Access the model
Visit Hugging Face, LEAP, or Liquid AI's Playground to access LFM models. Choose the appropriate model version based on your hardware and use case requirements.
-
Set up environment
Install required dependencies for your target hardware platform (NVIDIA, AMD, Qualcomm, etc.). Verify compatibility with your device specifications before proceeding.
-
Load the model
Download the model weights from the chosen platform. For on-device use, ensure your device meets the 900 MB memory requirement for LFM2.5-1.2B-Thinking.
-
Configure inference
Set up the model parameters for your specific task. Adjust temperature and max tokens for generation tasks, or enable thinking traces for reasoning-heavy applications.
-
Deploy application
Integrate the model into your application workflow. For enterprise deployments, contact Liquid AI for hardware-specific optimization guidance and scaling support.
Frequently Asked Questions
What is the Liquid Foundation Model (LFM) in AI?
LFM is a class of efficient generative AI models developed by Liquid AI for enterprises. Built using dynamical systems and signal processing principles, these models handle sequential data like text and video. They range from 1B to 40B parameters, optimized for edge devices to data centers.
How does LFM2.5-1.2B-Thinking work on mobile devices?
LFM2.5-1.2B-Thinking operates entirely on phones with just 900MB memory, enabling offline AI reasoning. It generates thinking traces before answers, enhancing performance on complex tasks. The model outperforms larger alternatives like Qwen3-1.7B in benchmarks while maintaining minimal memory usage.
What industries benefit most from LFM AI models?
LFM targets sectors needing efficient, private AI like financial services, biotech, and consumer electronics. Its edge deployment suits mobile apps, embedded systems, and on-premise solutions. The models' small footprint and hardware optimization make them ideal for resource-constrained environments.
How does LFM-3B compare to larger AI models?
LFM-3B delivers performance comparable to 7B-13B parameter models while being significantly smaller. It matches Phi-3.5-mini capabilities despite being 18.4% smaller. This efficiency comes from its dynamical systems architecture, making it suitable for mobile and edge applications.
Where can developers access LFM models?
LFM models are available through Hugging Face, LEAP, and Liquid AI's Playground. Partnerships with Qualcomm, Ollama, and FastFlowLM enable diverse deployment scenarios. The models support private, edge, and on-premise implementations across optimized hardware platforms.
What are the limitations of LFM AI models?
LFM requires custom enterprise pricing quotes with no transparent tiers. Performance gains vary across benchmarks compared to Llama or Gemma. Deployment needs hardware-specific optimization, and architectural details lack public documentation for independent verification.
Alternatives
How LFM (Liquid Foundation Model) compares
Direct head-to-head against 2 competitors. Picked by 7wData.
LFM (Liquid Foundation Model)
- Pricing
- Custom quote for enterprise solutions
- Target
- Liquid Foundation Models (LFM) are a new class of generative AI models developed by Liquid AI, designed for enterprises and developers needing efficient, scalable AI
- Strength
- LFM2.5-1.2B-Thinking outperforms Qwen3-1.7B in reasoning benchmarks while using less memory.
- Watch for
- Enterprise pricing requires custom quotes, lacking transparent tiered plans for smaller teams.
OpenELM
- Pricing
- Open-source (Apache 2.0)
- Target
- Edge devices, Apple ecosystem
- Deployment
- On-device, cloud
- Strength
- Optimized for Apple hardware
- Watch for
- Limited context window (1k tokens)
Phi-1.5
- Pricing
- Open-source (MIT)
- Target
- Resource-constrained environments
- Deployment
- Cloud, edge
- Strength
- Strong reasoning at small scale
- Watch for
- Narrow focus on synthetic data
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.