Liquid AI
Non-Transformer AI models built for efficient on-device inference
Overview
Liquid AI, an MIT CSAIL spinout, builds non-Transformer Liquid Foundation Models (LFM2.5) for on-device and edge inference — open weights on Hugging Face, free under $10M revenue, with an enterprise license and the LEAP SDK for iOS/Android deployment above that threshold.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Full commercial use of open LFMs
- Fine-tune & deploy anywhere
- No copyleft restrictions
- Custom licensing
- OEM & on-prem deployment
- Dedicated SLA support
Use cases
What you can produce with Liquid AI
- A production-ready on-device AI model integrated into an iOS or Android app via the LEAP SDK
- A fine-tuned domain-specific LFM for structured data extraction or RAG deployed on local enterprise hardware
- A real-time multilingual translation pipeline running fully offline on a smartphone without cloud dependency
- A fraud detection scoring service powered by LFM that processes transactions faster than a baseline transformer model
- A task-specific Nano model for data extraction or tool-calling deployed in an agentic workflow on consumer hardware
- A vision-language model embedded in an automotive HMI system, benchmarked smaller and faster than the prior baseline
- Open-weight LFM2 weights customised via fine-tuning and served through a self-hosted vLLM or llama.cpp endpoint
ASEAN Perspective
Liquid AI in Southeast Asia
Liquid AI has no APAC data centres, Asia-Pacific offices, or publicly named Southeast Asian enterprise customers as of June 2026, making it a challenging choice for ASEAN organisations subject to Singapore's PDPA, Malaysia's PDPA, Indonesia's PDP Law, or sector-specific MAS/OJK data-residency mandates. The core architectural advantage — highly efficient on-device inference — is, however, well-aligned with ASEAN's growing base of smartphone manufacturers, automotive OEMs (Thailand, Indonesia), and fintech firms that need low-latency AI without reliance on US or European cloud infrastructure. The LFM2.5-1.2B Japanese-English translation Nano and multilingual support (Chinese, Arabic, Korean) signal early interest in East Asian markets, and the edge-first architecture could enable air-gapped or locally-hosted deployments that satisfy data-sovereignty requirements without a dedicated regional cloud footprint — provided enterprises are prepared to self-host.
Liquid AI's bet is architectural: skip Transformers, build on liquid neural networks instead, and win on efficiency rather than raw scale. The numbers back it up so far — LFM2 claims 200% higher CPU throughput than Qwen3 or Gemma 3n, and a payment-processor case study credits the model with catching roughly $230M more fraud annually than the legacy system it replaced. LFM2.5 (January 2026) adds on-device agentic variants across ten-plus languages, including Japanese, Korean and Chinese. What's missing: a self-service hosted API. Access runs through the Playground, Hugging Face, or routers like OpenRouter, more friction than OpenAI or Mistral require. Enterprise pricing beyond the $10M-revenue free threshold is sales-gated and undisclosed. Named customers skew US and Gulf (a G42 partnership, not Asia); no APAC infrastructure or Asian enterprise logos have surfaced publicly as of mid-2026.
What people say
No self-service API, no published enterprise price list — Liquid AI still makes you talk to sales before you can deploy at scale, and that's the biggest friction point developers raise about an otherwise technically distinctive product. The company, spun out of MIT CSAIL in 2023 by Ramin Hasani, Mathias Lechner, Alexander Amini and Daniela Rus, skipped the Transformer architecture that basically every other foundation-model company builds on, betting instead on liquid neural networks — a time-continuous recurrent design meant to run faster and lighter on CPUs and edge hardware.
That bet has produced real numbers: LFM2 claims roughly double the CPU throughput of Qwen3 or Gemma 3n at comparable sizes, and Liquid's own case study for a payment processor puts the fraud-detection uplift at about $230M annually versus the prior model — a strong claim, published by Liquid itself rather than independently verified. LFM2.5 (January 2026) extends the lineup with on-device agentic models across ten-plus languages, and the LEAP SDK gets iOS and Android developers running models in about ten lines of code.
Commercially, everything under $10M in company revenue is free, including fine-tuning and on-prem deployment — a genuinely generous threshold. Cross that line and pricing turns custom and sales-gated. Recent partnerships lean US and Middle East (G42 in Abu Dhabi, a German reseller); there's no visible APAC infrastructure or Asian enterprise customer story yet, which matters for any regional buyer evaluating data residency.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- Liquid neural networks were originally inspired by the nervous system of the microscopic worm C. elegans, which navigates its environment with just 302 neurons.
- The LFM2 family was benchmarked at 200% higher throughput than Qwen3 and Gemma 3n on a standard CPU — without a GPU.
- Liquid AI's $250M Series A (December 2025, led by AMD Ventures) valued the two-year-old startup at $2.35 billion.
- A global automaker integrated Liquid's vision-language model and shipped a 50% smaller, 10x faster deployment in under one week using LEAP.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Liquid AI's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Liquid AI unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Liquid AI directly →
Spotted something out of date? Suggest an update →
Liquid AI in the news
More in LLMs & Chat
