Device-native model architecture
Liquid Foundation Models are built for efficiency and real-world constraints, with the site describing a proprietary architecture focused on speed, memory efficiency, and deployment on any device.
Liquid AI
ПодтвердитьLiquid AI builds foundation models for on-device AI across phones, cars, robots, and edge hardware. Open models and LEAP deployment workflow available.
Overview
Liquid AI builds foundation models designed to run natively on device and across edge environments, rather than only in data centers. The company positions these Liquid Foundation Models as a fit for phones, laptops, cars, robots, wearables, and embedded systems where latency, privacy, and hardware constraints matter.
The product surface includes open model families, the LEAP SDK for specialization and deployment, and a pricing and licensing model that allows free commercial use until a company exceeds $10 million in annual revenue. Liquid also publishes research through Liquid Labs, and the site names deployment targets and runtimes such as on-device, edge, cloud, llama.cpp, MLX, ONNX, CoreML, SGLang, and vLLM.
Core capabilities
Multiple model families and modalities
Models are available across text, vision-language, audio, and nano model families, giving teams a range of sizes and modalities for different workloads.
Broad hardware and runtime support
The models page says LFMs run on CPUs, GPUs, and NPUs, and the homepage names runtimes such as llama.cpp, MLX, ONNX, CoreML, SGLang, and vLLM.
Fine-tuning and deployment workflow
Liquid says its models are designed for rapid customization and fine-tuning, and LEAP lets users fine-tune a model, bake it to a runtime, and ship it.
Open commercial usage terms
The pricing page states that open models are free to download, run, and fine-tune, including commercially, until a company passes $10 million in annual revenue.
Public research foundation
Research is published openly through Liquid Labs, with papers, benchmarks, and technical reports tied to the model family.
Where it fits
On-device assistants and local inference
Use LFMs when you need AI that can run locally on phones, laptops, or embedded hardware and still respect latency and privacy constraints.
Task-specific model adaptation
Use the model families and customization workflow to adapt a foundation model for a specific task, then deploy it to a target runtime through LEAP.
Multimodal product experiences
Use the multimodal and low-latency model families for applications that combine text with images, or audio with text, in edge-aware products.
Embedded and OEM deployments
Use Liquid’s published deployment targets and commercial licensing model when building products for vehicles, OEM systems, or other edge deployments that may later need enterprise support.
Research and prototyping
Use the open research and model releases as a base for experimentation, benchmarking, and internal evaluation before moving to production.
Pros and Cons
Pros
- Runs across device classes including phones, laptops, cars, robots, and embedded hardware.
- Supports multiple runtimes and deployment targets, including CPU, GPU, NPU, on-device, edge, cloud, and on-prem scenarios.
- Offers a fine-tune-and-ship workflow through LEAP for moving models from notebook to production.
- Lets teams use open models commercially at no cost under the published revenue threshold.
- Keeps fine-tuned models private because the license has no copyleft requirement.
- Publishes research alongside the product, which helps explain the architecture and design direction.
Cons
- The source does not provide a full public docs set on the pages reviewed, so some implementation details remain unclear.
- Commercial use is free only up to the stated $10 million annual revenue threshold; larger companies need an enterprise license.
- Model and deployment choices are broad, but the pages reviewed do not spell out every supported framework, runtime, or workflow in detail.
FAQ
Is Liquid free to use commercially?
Liquid’s open foundation models are free to download, run, and fine-tune, including in commercial products, until a company passes $10 million in annual revenue. Above that threshold, a commercial license is required.
Will Liquid models run on my hardware?
Yes. The pricing page says the models can run on CPUs, GPUs, and NPUs across phones, laptops, vehicles, and embedded devices, with model sizes ranging from a few hundred million to a few billion parameters.
Do I have to release my fine-tunes?
The pricing page says there is no copyleft requirement, so fine-tuned models can remain private and proprietary.
What is the path for enterprise teams?
Liquid supports both self-serve use through open models and a contact-sales path for enterprise licensing, OEM and on-prem deployment support, and dedicated support.
How does the deployment workflow work?
The source says LEAP is the fastest path from notebook to production: fine-tune a model, bake it to a runtime, and ship it. The models page also notes support for on-device, edge, and cloud deployment.
Quick Facts
- Category
- Foundation models / developer platform
- Primary users
- Developers, researchers, startups, and enterprise teams
- Deployment targets
- Phones, laptops, cars, robots, wearables, embedded devices, edge, cloud, and on-prem
- Runtimes mentioned
- llama.cpp, MLX, ONNX, CoreML, SGLang, vLLM
- Commercial terms
- Free until a company exceeds $10M in annual revenue
- Source domain
- liquid.ai