d1-2B

About

d1-2B is Liquid AI's open decision model for text, pictures, and video, built on LFM3-VL-2B. Unlike generative LLMs it returns calibrated answers in one forward pass without generating tokens: typed answers (yes/no probabilities, choices, scores, text/number/date with span, object boxes, quantity intervals) for any number of named questions about one state, e.g. via model.system_one(state, questions, media=[...]).

• Best-in-class: leads the open d1 models on average and on Liquid AI's board of over a hundred decision benchmarks, ahead of d1-3B at under two thirds of its size.
• Beyond classification: extracts text, numbers, and dates with their span, locates objects in pictures, and estimates quantities with a p10–p90 interval.
• Millisecond decisions: 5.7 ms per decision on an NVIDIA H100 PCIe and 3.8 ms on an AMD MI325X (bf16, warm calls after model.compile()), with any number of questions over one state in one call.

Architecture (config.json, model type lfm3_vl, class D1Model): LFM3 tri-hybrid backbone, 24 layers — 13 gated short convolutions, 5 sliding-window attention (window 256) and 6 global attention layers (GQA 16/4 KV heads, head dim 128, partial RoPE factor 0.5, sandwich norm, tied embeddings) — with a 14-layer LFM3 vision encoder and GELU projector (hidden size 4096). 1.92B parameters in BF16, 128,000-token vocabulary, 128K-token context. Questions are answered as a set in one pass (d1_set_size 16 by default); input media are pictures and video.

Usage: input is a state (text or JSON), pictures or video, and any number of named questions; answer types are noul (yes/no), choice, score, text, number, date, box, boxes. Runs in Transformers >= 5.19 with its own model code (AutoModel.from_pretrained("LiquidAI/d1-2B", trust_remote_code=True)); there are no sampling parameters since nothing is generated. It is not a chat model — designed for pipeline steps that need a decision: routing and triage, moderation and guardrails, intent and topic classification, extraction from documents and forms, judging and reranking, visual inspection. Answers get less reliable past about 64K tokens of state.

Performance (model card): Decision Index public suite (v0.3) 51.1, DecisionBench (Hanno Labs) v1 77.1, Decision Bench (decisionbench.ai) v4 85.2, JevBench v1.4.1 77.9, ImajevBench v2.0-lite 75.5, WorkflowEvals 81.6; vision — AI2D 76.7, CV-Bench 86.1, MMBench 87.9, MME 85.4, MMMU-Pro 26.4, MMStar 67.5, POPE 88.7, VisualWebBench 96.2 (average 70.2 across the public benchmark set shared by all open d1 models).

Languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish. Repo published on Hugging Face on 2026-10-24 under the LFM Open License v1.0 (commercial use conditioned on ≤ USD 10M annual revenue, §5).

Training Data Post-trained decision model built on LFM3-VL-2B (LFM3 family). No public training-data details on the model card.