by Liquid·Released 25 February 2026
LFM2-24B-A2B is the largest model in the LFM2 family of hybrid architectures designed for efficient on-device deployment. Built as a 24B parameter Mixture-of-Experts model with only 2B active parameters per...
Reads about 120 pages
Standard OpenAI fields, grouped by what they control. Anything not listed is accepted and ignored.
Shaping the reply
How long, how varied, where to stop.
max_tokenstemperaturetop_ptop_kmin_pstopfrequency_penaltypresence_penaltyrepetition_penaltylogit_biasStructured output
Get JSON you can parse instead of prose.
response_formatNo. Liquid has retired it. Other Liquid models are on sale.
The lab has not said.
Liquid names no languages for LFM2-24B-A2B. Arabic, French and Darja: not stated.
It reads text. Up to 128K tokens in one request, about 120 pages. Longest reply: not stated.
Yes, its weights are public. Open weights on Hugging Face
Liquid. Released 25 February 2026.
Liquid d1 takes a situation described in text and pictures, plus questions of a fixed kind, and answers each with a yes or no, a score or a pick from options, with a probability and never with written text. Use it to route requests, sort items and run policy checks where you want a fast, scored answer to act on. It reads about 65 pages at once.
Liquid d1-ultra is the largest decision model from Liquid AI. You send a situation made of words and pictures with questions of a fixed kind, and for each one you get a yes or no, a score or a pick from a list, with a probability and no written text. It suits moderation, routing, sorting and checks on pictures.
This model accepts both text and images to understand a current situation. It can process up to 65 pages of information and generate up to 58 pages of structured choices. Its primary purpose is to act as a decision-making system, taking a defined state and specific questions to output a clear choice. Because it is built for structured outcomes, it is not suited for tasks requiring open-ended creative generation.
LFM2.5-24B-A2B is a Liquid AI model that thinks before it answers. You send text along with descriptions of your tools, and it thinks first, then either answers or uses a tool. It reads about 128 pages and writes up to 8 pages, so use it for automated helpers that use tools. Liquid recommends giving it document search or web search on questions that rest on facts, so pick a larger model when you have no search to offer.
From what its maker and two benchmark sites publish, and where it stands on OpenDunes. Read 7 Nov 2026.
Thinks first No.
Works through the problem before it answers.
Uses your tools No.
Can call functions your code provides.
Fills your JSON No.
Answers in the JSON shape you give it.
Searches the web No.
Can look things up online while it answers.
Cheaper on repeats No.
A prompt it saw a moment ago costs less the second time.
Open weights
The model's weights are public.