Liquid AI
Liquid AI is an artificial intelligence research and applications company founded by researchers from the Massachusetts Institute of Technology (MIT). The company focuses on developing next-generation foundation models based on Liquid Neural Network architectures, with core technology derived from research into dynamical systems and neural differential equations. Compared to traditional Transformer architectures, Liquid AI's models are designed to perform inference and sequence modeling tasks with greater parameter efficiency and computational efficiency, targeting enterprise AI applications.
Core Facts
Decision models do not generate text; instead, given an input state and a set of named questions, they return typed answers with probabilities in a single forward pass.
Liquid AI's technical foundation is Liquid Neural Networks, inspired by research on the nervous system of the nematode C. elegans.
Liquid AI has expanded its D1 decision model with visual capabilities, enabling it to accept images, text, or both, and return probabilities for yes/no judgments, multiple-choice, or scoring questions.
Timeline (last 90 days)
LFM3 is a model series launched by Liquid AI, and its lfm3-audio and lfm3-omni variants support text-to-speech capability.
LFM is Liquid AI's foundation model product line, positioned as a lightweight, efficient, locally deployable general-purpose model family; LFM3 is the third-generation version of the series, with significantly expanded multimodal capabilities compared with its predecessors.
The COO of Liquid AI is Jeffrey Li.
The six LFM3 models use open weights, share the same backbone network, and cover text, vision, audio, omni-modal and decision models.
Currently the public information focuses on model specifications and deployment conditions, and the official side has not yet given the specific capability boundaries of each modality, actual inference speed, or horizontally comparable data on performance across different hardware.
Liquid AI positions LFM3 for three types of edge devices — phones, laptops and robots — and the goal is not cloud inference.
The 1.7B text model can be fine-tuned on a single GPU.
LFM3-1.7B adopts a tri-hybrid backbone structure, supports a 256K context window, and is about 1GB in size under 4-bit quantization.
The six models in the LFM3 series cover directions including text, vision, audio, omni, and decisions.
Official data shows that on a single B200, LFM2.5-1.2B's p95 latency on RIPL is 14.4 to 28.4 times lower than Qwen3.5-0.8B's on vLLM.
40 more timeline events
All Facts (20)
Decision models do not generate text; instead, given an input state and a set of named questions, they return typed answers with probabilities in a single forward pass.
65%VerifiedLiquid AI's technical foundation is Liquid Neural Networks, inspired by research on the nervous system of the nematode C. elegans.
65%VerifiedLiquid AI has expanded its D1 decision model with visual capabilities, enabling it to accept images, text, or both, and return probabilities for yes/no judgments, multiple-choice, or scoring questions.
70%UnverifiedLiquid AI由MIT CSAIL的Ramin Hasani、Mathias Lechner等研究人员于2023年创立
85%UnverifiedLFM is Liquid AI's foundation model product line, positioned as a lightweight, efficient, locally deployable general-purpose model family; LFM3 is the third-generation version of the series, with significantly expanded multimodal capabilities compared with its predecessors.
50%UnverifiedLFM3 is a model series launched by Liquid AI, and its lfm3-audio and lfm3-omni variants support text-to-speech capability.
50%UnverifiedThe COO of Liquid AI is Jeffrey Li.
50%UnverifiedCurrently the public information focuses on model specifications and deployment conditions, and the official side has not yet given the specific capability boundaries of each modality, actual inference speed, or horizontally comparable data on performance across different hardware.
50%UnverifiedThe 1.7B text model can be fine-tuned on a single GPU.
50%UnverifiedThe six LFM3 models use open weights, share the same backbone network, and cover text, vision, audio, omni-modal and decision models.
50%UnverifiedLiquid AI positions LFM3 for three types of edge devices — phones, laptops and robots — and the goal is not cloud inference.
50%UnverifiedLFM3-1.7B adopts a tri-hybrid backbone structure, supports a 256K context window, and is about 1GB in size under 4-bit quantization.
50%UnverifiedThe six models in the LFM3 series cover directions including text, vision, audio, omni, and decisions.
50%UnverifiedLiquid AI used a set of comparative figures to push the "specialized inference engine" category back into the spotlight.
50%UnverifiedOfficial data shows that on a single B200, LFM2.5-1.2B's p95 latency on RIPL is 14.4 to 28.4 times lower than Qwen3.5-0.8B's on vLLM.
50%UnverifiedParticipants offering alternatives to decision models include platform-level players such as Cloudflare, OpenAI, Perplexity, and AWS Strands Labs, as well as Liquid AI and a group of independent developers.
50%Unverifiedd1-ultra is a multimodal decision model specifically handling classification, routing, and scoring tasks; its input supports text and images, and it does not generate text but rather outputs structured probabilistic judgments.
50%UnverifiedThe core capability emphasized officially by d1-ultra is calibrated probabilities, meaning the model's output probabilities should match actual occurrence frequencies.
50%UnverifiedData released by Liquid AI shows that LFM2.5-24B-A2B beats Qwen3.6-35B-A3B on tool-use benchmarks
50%UnverifiedModel vendors are beginning to optimize specifically for the on-device agent niche scenario rather than simply shrinking large models.
50%