LiquidAI D1 Decision Model
a day ago
- LiquidAI/d1-3B is a 3.1B parameter multimodal decision model that returns calibrated, typed answers (yes/no, choices, scores) in a single forward pass with zero output tokens.
- It supports text, JSON, and images as input and is designed for tasks like routing, moderation, intent classification, extraction checks, reranking, and visual inspection, not for chat.
- The model can be used via Transformers (with trust_remote_code), vLLM, SGLang, Docker Model Runner, and in notebooks such as Google Colab and Kaggle.
- It achieves a Decision Index score of 48.57, outperforming many larger models, and offers fast inference (8 ms per decision on an NVIDIA RTX 4090).
- Open source access requires transformers>=5.14 and trust_remote_code=True; the model is available on Hugging Face and can be deployed locally or on edge devices.