Liquid AI

Build efficient general-purpose AI at every scale.

[Liquid AI @liquidai]

Today we release LFM2.5-2.6B, an agentic model that runs entirely on-device. It plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots. Data never leaves the device, and the marginal cost of each run is essentially zero.

> Pre-trained on ~34T tokens
> LFM2.5 flagship hybrid architecture
> Context length: 128K
> Vocab size: 128K
> balanced intelligence per watt
> customizable on a single GPU for any specialized task
> LFM2 open-weight license

Comparable or better scores compared to models up to nearly 4x its size:
> ToolSandbox 77.83, ahead of Qwen3.5-9B at 76.44
> Multi-IF 80.07, ahead of Gemma-4-E4B-it at 77.35
> IFStruct 85.49, ahead of Qwen3.5-9B at 78.50

🧵

[Ramin @ramin_m_h]

With Insilico, we got our LFM 2.6B model customized on retrosynthesis and it performs better than Astra! 🌊🌊🌊

[Ryan Wright @IQReactorAI]

Guys and Girl nerds, the tiny 500m-1.2B models trained for a single task are taking off cause they are dirt cheap to run and do that task better then frontier models.

Us Local AI folks can run AND train these at home. (I will soon)

Actually ☝️🤓 For a training a highly specialized 1B model, 1,000 to 5,000 high-quality, varied examples are often enough to achieve frontier-level performance on that specific task. Not 50,000 to 1 million like with big models.

Checkout this blog article by @ liquidai where they talk about their tiny task models.

"LFM2-Extract – A 350M and 1.2B multilingual models for data extraction from unstructured text, like turning invoice emails into JSON objects.

LFM2‑350M‑ENJP‑MT –  A 350M model for bidirectional English ↔ Japanese translation.

LFM2‑1.2B‑RAG – A 1.2B model optimized for long‑context question answering in RAG pipelines.

LFM2‑1.2B‑Tool – A 1.2B model built for function calling and agentic tool use.

LFM2‑350M‑Math –  A 350M reasoning model for solving mathematical problems.

Luth-LFM2 – An additional community-driven series of French fine-tunes to enable general-purpose assistants for on-device chat."

Read more here: https://t.co/GqsOAhUD3C

[Vasko @RoliumGens]

Introducing Cyber-Prime 1 2.6B.

It's the smallest cyber-security agentic model you can find, built on top of @liquidai's LFM 2.6B ( goated model).

It has been trained on many reasoning traces, cyber-security dataset and synthetic data I have built.

I have also put this model inside of my recursive self improvement loop. As this is a smaller model, new checkpoints are going to come extremely quickly ( first one as soon as tomorrow)!

A 4B version is coming soon as well as a 7.9B MoE so stay tuned!

🤗 Weights : https://t.co/MGzn5Yp7cd

[Liquid AI @liquidai]

Can a language model master chemical synthesis through deep customization? Our partners @InSilicoMeds tested this out with our LFM2.5-2.6B.

See what happened next 👇

[Insilico Medicine @InSilicoMeds]

Can a language model become SOTA in chemical synthesis? 🧪🔥

We put a 2.6B-parameter @liquidai model into MMAI Gym.

It now sets the state of the art in single-step retrosynthesis. 🚀

#insilicoSOTAFM

[Ivan Fioravanti ᯅ @ivanfioravanti]

Told you that @liquidai LFM2.5-2.6B and even LFM2.5-VL-3B are SoTA. This lab rocks!

[Bogdan A. Zagribelnyy, PhD @sumrexromanus]

It was a pleasure to work with #LFM by @liquidai @ramin_m_h to train our new #insilicoSOTAFM model for single-step retrosynthesis 🧪on ~46M reactions! It provides plenty of diverse reactions that only partially intersect with the reaction space by previous SOTA small models! 🔎👀

[Insilico Medicine @InSilicoMeds]

Can a language model become SOTA in chemical synthesis? 🧪🔥

We put a 2.6B-parameter @liquidai model into MMAI Gym.

It now sets the state of the art in single-step retrosynthesis. 🚀

#insilicoSOTAFM

[Liquid AI @liquidai]

👇

[Ramin @ramin_m_h]

about a year ago, we released the first instances of Liquid nanos. these are products we sell to enterprises: tiny models + the customization platform matching the quality of frontier models on specialized use-cases. It is easy to recursively customize a model to 90% production quality. going from 90 to 100 that satisfies enterprise reliability, and QA is the art. It is very hard, but joyfully doable

Instances of these Liquid nanos are today in full production at @Shopify at scale!

https://t.co/LWj3uSS6Li

[Viviana Márquez @vivmarquez]

LFM2.5-2.6B is a compact reasoning model with just 2.6B parameters: small enough to run on a phone and specifically trained for agentic workflows (planning, calling tools, and tackling multi-step tasks.)

Its post-training included agentic reinforcement learning inside real agent harnesses such as OpenClaw and Hermes Agent.

If you're already using OpenRouter, trying it is as simple as changing the model slug: liquid/lfm-2.5-2.6b:free

Give it a try: https://t.co/fGndbYNuMw

[Ramin @ramin_m_h]

Proud to see my co-founder, phd co-advisor and postdoc advisor Daniela Rus, on the TIME's 100 most influential people in AI.

Daniela is one of the defining figures of modern robotics and AI. it is a humbling experience to share a 9 years professional journey with her as a mentor, advisor, and business partner.

As she describes it herself about her work @MIT_CSAIL and @liquidai, it is "all about the science & engineering of intelligence, & especially embodied intelligence” ✨

https://t.co/sdoeOtigNc

[Ramin @ramin_m_h]

coming to Pipette!
https://t.co/dUdBzo41nh

[Abdur Rahim @_ARahim_]

Wow! 387 tok/s on my M4 Pro! This is LFM2.5-1.2B-instruct in BF16 with mlx-dspark.

https://t.co/eju6FLm7SM

@liquidai

[Viviana Márquez @vivmarquez]

You can now benchmark how a model actually performs on the device you plan to deploy it to, before you ship.

I tried Pipette by @liquidai on my iPhone 13 Pro Max with four instruct models, and the results made the tradeoffs pretty clear 🧵

[Adrien Grondin @adrgrondin]

Incredible release!

About time we had something easily accessible like this.

Choose the best model for your needs and iPhone (data only for the 17 Pro for now) then you can try it out in @LocallyAIApp.

[Viviana Márquez @vivmarquez]

In case you've wondered what model to run on the device you probably use the most

[Mathias Lechner @mlech26l]

This will be a game changer for local AI

[Maxime Labonne @maximelabonne]

Great to see on-device evaluations!

Very proud of LFM2.5-2.6B. The way it generalizes beyond agentic tasks is crazy.

[Mikhail Parakhin @MParakhin]

Liquid AI's models are the best for the scaled production deployment - pareto-optimal, beating much larger rivals. That's why we use them at Shopify.

[Vishnu Samavedula @svrvishnu]

Looking to implement intelligence on edge? Check out @ArtificialAnlys 's new testing results.

[Leonie @helloiamleonie]

Which is the best model to run on an iPhone 17 Pro?

Now you'll know:

Our team at Liquid AI has partnered with Artificial Analysis to bring you an open-source benchmark to measure on-device quality, speed, latency, and memory.