Liquid d1-3B

Liquid d1-3B is Liquid AI's 3B-class multimodal decision model. Its official Hugging Face card identifies LFM2.5-VL-3B as the base model and labels the intended work as decision, classification, calibration, edge, and image-text-to-text inference. It is a long-tail option for teams that need a smaller, task-oriented model rather than a general chat model.

Where it fits

Use d1-3B to make bounded choices from text and images: routing requests, classifying documents, applying a policy label, or deciding whether a workflow should escalate to a larger model. Its card lists a multilingual language set including English, Chinese, and Japanese. The published weights use Liquid AI's lfm1.0 license, so review that license before redistribution or production deployment.

Practical workflow

Start with a labelled sample from the exact decision you need to make. Define a small output schema, test error cases and ambiguous inputs, then measure accuracy and calibration separately. Route low-confidence or high-impact cases to human review or a stronger model. This is particularly useful when a predictable decision path matters more than open-ended writing.

Limits and alternatives

The model card does not make it a replacement for a frontier general-purpose model. Treat its small size and specialist framing as reasons to evaluate carefully on your images, languages, and policy edge cases. For local agent work with native tools and long context, compare LFM 2.5-2.6B. For broad multimodal retrieval, compare EmbeddingGemma.

Getting started

Read the official model card, verify the current files and license, and run a held-out evaluation before connecting the model to automated decisions.