Liquid Foundation Model

Liquid Foundation Models (LFMs) are a new generation of generative AI models developed by Liquid AI, a company founded in 2023 as an offshoot from MIT.

Updated 8 Jun 2026 Reviewed by 7wData

Publisher review

Liquid Foundation Models (LFMs) are a new generation of generative AI models developed by Liquid AI, a company founded in 2023 as an offshoot from MIT. Unlike the vast majority of large language models that rely on the Transformer architecture, LFMs are built from first principles using a non-Transformer-based design. This architectural shift is aimed at organizations and developers who need state-of-the-art language model performance but face constraints on memory, latency, or hardware — particularly for deployment on edge devices, embedded systems, or processors outside of data centers.

The first series of LFMs includes three sizes: 1 billion, 3 billion, and 40 billion parameters, making them suitable for a range of applications from lightweight on-device inference to more demanding enterprise tasks. Liquid AI positions these models for use cases where traditional Transformer models are too memory-intensive or slow, such as real-time inference on consumer hardware or in-car intelligence, as evidenced by their partnership with Mercedes-Benz to scale embedded in-car intelligence.

How it works

Non-Transformer Architecture

Built from first principles without relying on the Transformer architecture, offering a fundamentally different approach to sequence modeling.
Long Context Length

Supports a context window of up to 32,000 tokens, enabling processing of lengthy documents or conversations in a single pass.
Smaller Memory Footprint

Requires significantly less memory than comparable Transformer models, as reported in community benchmarks and official documentation.
Efficient Inference

Designed for fast inference on a variety of devices, including consumer GPUs and edge hardware, without sacrificing output quality.
Multi-Scale Model Series

Offered in 1B, 3B, and 40B parameter sizes, allowing users to match model capacity to their compute budget and latency requirements.
Hardware Optimization

Optimized for NVIDIA, AMD, Qualcomm, Cerebras, and Apple hardware, ensuring broad compatibility across different deployment environments.
State-of-the-Art Benchmarks

Achieves competitive or superior performance on many standard NLP benchmarks relative to Transformer models of similar scale.

Strengths and trade-offs

Strengths

Achieves state-of-the-art performance on multiple NLP benchmarks while using a non-Transformer architecture that reduces memory consumption by an unspecified but significant margin.
Supports a 32,000-token context window, enabling analysis of long documents or extended conversations without truncation.
Offers efficient inference that allows the 1B and 3B models to run on consumer-grade hardware, including laptops and edge devices.
Built from first principles, providing a novel architectural approach that may offer advantages in latency and hardware flexibility over traditional Transformer models.

Trade-offs

Limited independent third-party benchmarks and real-world deployment reports are available, making it difficult to verify performance claims outside of Liquid AI's own testing.
The non-Transformer architecture may require custom inference stacks or specialized libraries, potentially increasing integration effort for teams accustomed to standard frameworks.
Pricing and licensing details are not publicly specified, creating uncertainty for enterprises evaluating total cost of ownership.
The model series is early-stage (first announced in 2024), so the ecosystem of tools, fine-tuning recipes, and community support is less mature than for established models like GPT-4 or Claude.

Pricing context

Not publicly specified in available sources; likely enterprise-negotiated or usage-based, with no free tier or per-token pricing disclosed.

Getting started with Liquid Foundation Model

Sign up for Liquid AI

Go to the Liquid AI website and create an account. Provide your email and set a password, or use a supported single sign-on method. Verify your email to activate the account.
Access the model API

Log into your Liquid AI dashboard and navigate to the API keys section. Generate a new API key and copy it. Store the key securely, as it will be used to authenticate requests.
Choose a model size

Select the Liquid Foundation Model size that fits your compute budget and latency needs: 1B, 3B, or 40B parameters. Refer to the documentation for performance benchmarks on your target hardware.
Send your first inference request

Use your API key to send a POST request to the inference endpoint with a prompt. Include the chosen model ID and parameters like max_tokens. Parse the JSON response to get the generated text.
Deploy on your hardware

Download the model weights from Liquid AI's portal for the size you need. Load them into your application using the provided inference library. Optimize for your specific GPU or edge device following the hardware guide.

Frequently Asked Questions

What is a Liquid Foundation Model?

A Liquid Foundation Model (LFM) is a generative AI model from Liquid AI, founded in 2023 as an MIT offshoot. Unlike most large language models, LFMs use a non-Transformer architecture built from first principles, designed for efficient inference on edge devices and consumer hardware.

How is Liquid Foundation Model different from Transformer models?

Liquid Foundation Models are built from first principles without relying on the Transformer architecture. This design reduces memory consumption and enables faster inference on diverse hardware, including edge devices and consumer GPUs, while achieving competitive performance on NLP benchmarks.

What sizes are available for Liquid Foundation Models?

The first series of Liquid Foundation Models includes three parameter sizes: 1 billion, 3 billion, and 40 billion. This range allows users to select a model that fits their compute budget and latency needs, from lightweight on-device inference to demanding enterprise tasks.

Can Liquid Foundation Models run on edge devices?

Yes, Liquid Foundation Models are optimized for edge deployment. The 1B and 3B models run on consumer-grade hardware like laptops and edge devices, with support for NVIDIA, AMD, Qualcomm, Cerebras, and Apple hardware, making them suitable for real-time inference.

What is the context length of Liquid Foundation Models?

Liquid Foundation Models support a context window of up to 32,000 tokens. This allows processing of lengthy documents or extended conversations in a single pass without truncation, benefiting applications like document analysis and long-form dialogue.

What partnerships does Liquid AI have for Liquid Foundation Models?

Liquid AI has partnered with Mercedes-Benz to scale embedded in-car intelligence using Liquid Foundation Models. This partnership highlights the models' suitability for real-time, memory-efficient inference in automotive and other edge environments.

Alternatives

01-ai ↗
Aleph Alpha ↗
Cohere ↗

How Liquid Foundation Model compares

Direct head-to-head against 3 competitors. Picked by 7wData.

Pricing: Not publicly specified in available sources; likely enterprise-negotiated or usage-based, with no free tier or per-token pricing disclosed.
Target: Liquid Foundation Models (LFMs) are a new generation of generative AI models developed by Liquid AI, a company founded in 2023 as an offshoot from
Strength: Achieves state-of-the-art performance on multiple NLP benchmarks while using a non-Transformer architecture that reduces memory consumption by an unspecified but significant margin.
Watch for: Limited independent third-party benchmarks and real-world deployment reports are available, making it difficult to verify performance claims outside of Liquid AI's own testing.

Pricing: Custom/Contact sales
Target: Enterprise AI deployments with edge focus
Deployment: Cloud, on-prem, edge
Strength: Yi-34B model with 200K context window
Watch for: Limited public benchmarks vs. Liquid's task-specific models

Pricing: $1.50/M tokens for Luminous-base
Target: European enterprises needing sovereign AI
Deployment: Cloud or private servers
Strength: Multilingual Luminous models with explainability
Watch for: Higher latency than edge-optimized Liquid Nanos

Pricing: $15/M tokens for Command R+
Target: Business RAG and tool-calling workflows
Deployment: Cloud API
Strength: Strong retrieval-augmented generation pipelines
Watch for: No sub-1B parameter options for edge deployment

User reviews

No user reviews yet. Be the first to write one.

Sources

Reporting on this tool draws on these publicly available sources.

Publisher review

Get the AI & data signal, daily.

How it works

Non-Transformer Architecture

Long Context Length

Smaller Memory Footprint

Efficient Inference

Multi-Scale Model Series

Hardware Optimization

State-of-the-Art Benchmarks