Liquid AI’s d1-2B is a multimodal decision model, not a chatbot

Liquid AI’s d1-2B is a 1.92-billion-parameter model for making structured decisions from text, images and video, rather than generating chat. OrcaRouter says it leads the company’s open d1 models on its benchmark table, but that result has not been independently tested on d1-2B.

Watch Desk analysis

What happened

OrcaRouter describes d1-2B as a model that returns answers such as yes/no decisions, ranked choices, extracted values and object locations, with probabilities attached. It is based on Liquid AI’s LFM3-VL-2B and has a stated 128,000-token context window. The model card says answers become less reliable beyond about 64,000 tokens.

The report says the model’s repository appeared on 24 October and Liquid AI named it in a 27 October post about the LFM3 family. It also says MLX and ONNX versions are available, but there is no GGUF build or hosted API listed. Read OrcaRouter’s account of d1-2B.

Why it matters

This is aimed at the quieter plumbing of AI products: classifying, extracting or routing information inside a larger workflow. A model that returns a typed answer instead of composing a paragraph could suit those jobs, provided its results are reliable enough for the task.

OrcaRouter says Liquid AI’s benchmark table gives d1-2B a score of 51.1 against 49.0 for d1-3B on the same version of the Decision Index. That is the company’s result, not an independent d1-2B evaluation. The report also notes that d1-3B scores higher on some individual tests, including typed decisions and diagram understanding.

Our read

The interesting pitch is not “a smaller chatbot”. It is a compact model designed to make bounded choices, with multimodal input and outputs a software pipeline can use directly. That could be useful, but the benchmark lead is still a vendor-reported result, and the lack of a GGUF build or hosted endpoint narrows the ready-made deployment options. Teams considering it should test it against their own labelled examples before trusting its scores or probabilities.

What to watch

• Whether independent evaluators test d1-2B on the same benchmarks. – Whether Liquid AI publishes a GGUF build or hosted inference option. – How reliably it handles longer inputs and real-world tasks beyond the reported tests.

Discussion spark: For routine classification or extraction, would you trust a small specialised model with a vendor-reported benchmark lead, or wait for independent testing first?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.