Liquid AI
Liquid AI · Coding · Other
Liquid AI builds Liquid Foundation Models, a family of small open-weight AI models that run on phones, laptops and edge hardware without a cloud connection.
About Liquid AI
What Is Liquid AI
Liquid AI is a company that designs foundation models meant to run on the device in front of you, not in a distant data center. Its core product is the Liquid Foundation Models (LFM) family, a set of open-weight models small enough to fit on phones, laptops, wearables and embedded chips. The company came out of MIT's CSAIL lab in 2023 and has since shipped several model generations, with LFM2 and LFM2.5 as the current line.
The pitch is simple. Large cloud models charge per token and send every request to a server. LFM models run locally instead, so latency stays in the millisecond range, the work keeps going offline, and private data never leaves the machine. The models range from 350 million up to 8.3 billion parameters, which covers a text assistant on a budget phone and a reasoning agent on a laptop.
Two things shape how far you can take it. Licensing is free only until your company passes $10 million in annual revenue, after which you need a paid commercial license. And the smallest models trade raw fact recall for speed and size, so a 350M model will not match a large cloud model on open-ended knowledge questions.
Getting Started
- Pick a model on the Liquid AI site or the Hugging Face library and check the size against your target hardware. Variants run under 1GB for the smallest builds.
- Download the open weights, or pull them through your runtime of choice such as llama.cpp, MLX, ONNX, vLLM or SGLang.
- Run the model locally to see how it behaves on your device, or open the browser playground if you only want to test prompts.
- Fine-tune the model on your own data when a general model is not specific enough. Use the LEAP platform or open tools like Unsloth and TRL.
- Ship it inside your app, or in the cloud, depending on where your users are.
Product Information
A quick look at Liquid AI's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- App and mobile developers
- Privacy-focused teams
- Researchers and students
- Hardware makers
Tasks
- Building an offline chat assistant
- Adding image understanding
- Speech interaction
- Routing and classification
- Custom fine-tuning
Scenarios
- A vehicle assistant that answers in milliseconds and keeps working when the signal drops.
- A field app on a phone that summarizes documents and translates text with no cloud round trip.
- A regulated workflow where raw data stays on the machine and never leaves the building.
- A prototype that needs a working model on day one, before spending on cloud inference.
Key features
Open-weight model family
Liquid AI releases its models as open weights, from a few hundred million to a few billion parameters. You can download, run and fine-tune every model, including for commercial work, until your company passes $10 million in annual revenue. The smaller builds run under 1GB, which is small enough for a modern phone.
LFM2 and LFM2.5 generations
LFM2 covers dense models at 350M, 700M, 1.2B and 2.6B parameters plus a mixture-of-experts variant, all with 32K context. LFM2.5 adds Base, Instruct, Japanese, vision-language and audio-language models, with instruction following tuned for on-device agents. The newer text models double the vocabulary and reach up to 128K context.
Multimodal coverage
The vision models pair a lightweight text backbone with a SigLIP2 image encoder for on-device image understanding. The audio models handle speech in and out in one model instead of pairing a separate speech recognizer with a language model. That means a phone can read a photo or hold a spoken conversation without sending data away.
LEAP platform
LEAP, also known as the LEAP SDK, is Liquid's cross-platform toolkit for teams that build and deploy models on mobile and embedded devices. It gives iOS and Android a single workflow, works with Liquid's own models and other small open models, and includes a library with builds small enough for a phone's memory. LEAP Finetune handles data prep, model training and export in one place.
Runs on everyday hardware
LFMs run on CPUs, GPUs and NPUs, from phones and laptops to vehicles and embedded devices. They target millisecond latency for on-device resilience, so the model keeps answering even when the network is slow or gone. You choose the size that matches the chip you have.
Free to fine-tune
Fine-tuning AI models is not locked behind a paid tier. The license lets you train a model on your own data and keep the result private, with no copyleft requirement. Liquid's own fine-tuning tools sit alongside open options like Unsloth, TRL and the Transformers library.
Model Playground and Liquid Apollo
The Model Playground lets you test models in a browser before any download. Liquid Apollo is a mobile app that puts a local model in your pocket, a low-latency, cloud-free way to try on-device AI without setup.
Pros and cons
Pros
- No per-token bills: running locally removes recurring API costs, which matters for steady or high-volume work.
- Privacy by default: sensitive text and images stay on the device instead of going to a server.
- Small enough to ship: builds under 1GB fit phones, wearables and embedded chips.
- Free to start: every model is free to download and fine-tune until the $10 million revenue line.
- Genuine multimodal range: text, vision and audio models come from one family and one license.
Cons
- Revenue threshold: once your company passes $10 million a year, the free commercial license ends and you need a paid deal.
- Small models have limits: a 350M model will not match a large cloud model on open-ended knowledge or long reasoning chains.
- Hardware still matters: the faster and larger models need a capable CPU, GPU or NPU, so older devices get less out of them.
- Setup is developer work: getting a model into an app means handling runtimes and packaging, not just signing up for an account.
Frequently asked questions
Liquid AI is a company, spun out of MIT's CSAIL lab, that builds Liquid Foundation Models. These are small open-weight AI models designed to run on phones, laptops and edge devices rather than in a data center, so inference stays fast and private.
Related content
Explore related tools, skills, and articles for Liquid AI.
Liquid AI Alternatives
Forefront
Forefront · CodingForefront is a web platform for building with open-source AI. It lets you fine-tune leading open-source language models on your own data, evaluate how they perform, and run them through an API or export them to host yourself. Developers who want the convenience of a closed-source platform but insist on owning their models and data are the target audience here.
Startkit
StartKit.AI · CodingStartkit is a boilerplate for building AI SaaS and AI wrapper products. Think of it as an AI startup boilerplate with the boring parts already wired up: authentication, Stripe and Lemon Squeezy payments, usage limits, transactional email, and an AI API starter that talks to OpenAI, Anthropic, Groq, or Llama. You clone the repo, set your price, and start on the part of your product that people actually pay for. It's Next.js under React and Tailwind, so most of the boilerplate code already feels familiar.
Testim
Tricentis · CodingTestim is an AI-powered test automation platform for building and running end-to-end tests across web, mobile, and Salesforce applications. It leans on machine learning to keep tests stable when an interface changes, so teams spend less time fixing broken selectors. Not bad for an automated testing tool you can start using today. You create tests by recording actions in a browser, then optionally add JavaScript when you need more control. It's a solid pick for busy QA teams.