Liquid AI Releases d1-3B Model Weights on Hugging Face
Liquid AI has made the open weights for its d1-3B model available on Hugging Face for public access and evaluation. This release allows developers and researchers to run and benchmark the 3-billion parameter model on local hardware. The release provides the local LLM community with access to Liquid AI's novel non-transformer architectural design. It advances open-weights AI by giving developers an alternative approach optimized for structured decision-making and edge computing. The d1 architecture is designed as a zero-token decision model that evaluates text and visual inputs to output calibrated probabilities for classification, routing, and scoring in a single execution step rather than generating output text token-by-token. The model supports a 32k context window and multimodal inputs.
## BACKGROUND
Liquid AI is an artificial intelligence company focused on developing Liquid Foundation Models (LFMs), a class of architectures optimized for inference efficiency on edge devices and clouds. Traditional Large Language Models (LLMs) rely on autoregressive transformer architectures that generate output token-by-token, which can be computationally expensive for classification and decision-making tasks.