Skip to content

d1

3 TagsUpdated LFM Open License v1.0by LiquidAI

LiquidAI d1-3B: typed decisions with max-pooled answer forms, Q8_0.

multilingual3b
ollaya run d1 --preset triage "I was charged twice for my subscription this month and want a refund."

Models

View all
Name
d1latest2.9 GB · 32768 ctx · 100+ languages
d1:3b2.9 GB · 32768 ctx · 100+ languages
d1:3b-vision3.5 GB · 32768 ctx · 100+ languages

Readme

d1-3B by Liquid AI is a multimodal decision model based on LFM2.5-VL-3B. Ollaya reads its answer-token probabilities directly, with the author's prompt and max-pooled answer forms, and never generates an explanation.

Needs Ollaya 0.13.0 or newer, the first release that runs the d1-v1 layout. Update first (ollaya --version); an older version downloads the weights and then refuses to load them.

Models

TagWeightsImages
d1, d1:3bAuthor Q8_0 GGUF, 2.87 GBNo
d1:3b-visionSame GGUF plus the author's 583 MB Q8_0 projectorYes

Use the vision tag for image questions. Both tags use temperature 1.0, without additional calibration, and support predicate, choice and score questions. Scores take 2–10 levels. The context holds 32,768 positions. Questions run independently because the model has recurrent state.

Typed-decisions accuracy, the argmax against the majority label on all 400 states and measured by Ollaya on an RTX 5090, is 0.650 (choice 0.603, score 0.571, yes/no 0.802), with a calibration error (ECE) of 0.033 at temperature 1.0. On an RTX 4090 the runner answers five questions in about 48 ms; each question evaluates the whole state again, so long states with many questions take longer.

License

LFM Open License v1.0. Commercial use by an organization with US$10 million or more in annual revenue is not covered by the license and needs a separate agreement with Liquid AI. Weights and projector download directly from Liquid AI, pinned to a verified revision.