Liquid AI releases d1-2B, a 1.92B decision model that answers in one forward pass with no generated tokens

Liquid AI has released d1-2B, a 1.92B-parameter decision model for text, pictures and video that returns calibrated answers in one forward pass without generating tokens. The checkpoint is post-trained from LFM3-VL-2B and is published on Hugging Face under the LFM Open License v1.0. The model card is in English.

Specifications

• Parameters: 1.92B
• Architecture: LFM3 backbone, 24 layers: 13 gated short convolutions, 5 sliding-window attention, 6 global attention
• Context length: 128K tokens
• Vocabulary: 128,000 tokens
• Input: a state (text or JSON), pictures or a video, and any number of named questions
• Output: typed answers with their probabilities, with no generated tokens
• Languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish
• Pricing: not yet disclosed

The LFM3 family includes base, text, vision, audio, omni and decision models. d1-2B decides over text, pictures and video, while d1-Omni-2B adds audio.

How it works

d1-2B runs in Transformers (version 5.19 or later) with its own model code, so trust_remote_code=True is required. Because answers are read from a single forward pass, there are no sampling parameters to set. The main call is system_one(state, questions, media=None), which answers named questions over one state in one pass. system_one_batch handles many requests in one call. Each call returns answers plus a usage field that reports zero output tokens.

Question types include yes/no (returns P(yes)), single and multi-choice, ordered scores with 2 to 10 levels, text, number and date extraction with spans, numeric estimates with a p10 to p90 interval, and bounding boxes for one object or every instance (16 at most, or a set maximum). A value or box is returned as None when its found score is under 0.5. A set_size option controls how many questions are read together after the state. The default is the trained size of 16, and Liquid AI says larger values are faster but cost some accuracy on text.

Speed

Liquid AI claims 5.7 ms per decision on an NVIDIA H100 and 3.8 ms on an AMD MI325X, with any number of questions about one state in one call.

Benchmarks

According to Liquid AI, d1-2B leads the open d1 models on average and on most single benchmarks, ahead of d1-3B at under two thirds of its size. On Liquid AI's own board of over a hundred public and private benchmarks, d1-2B scores 63.2, against 61.8 for d1-Omni-2B and 58.8 for d1-3B.

Other decision results for d1-2B:

• Decision Index public suite (v0.3): 51.1
• DecisionBench (Hanno Labs) v1: 77.1
• Decision Bench (decisionbench.ai) v4: 85.2
• Fast Decisions: 58.8
• JevBench v1.4.1: 77.9
• ImajevBench v2.0-lite: 75.5
• WorkflowEvals: 81.6
• Typed Decisions: 61.3

On the vision suite, d1-2B has an average of 70.2. Individual scores include MMBench 87.9, CV-Bench 86.1, POPE 88.7, MME 85.4, VisualWebBench 96.2 and MMMU-Pro 26.4.

d1-2B does not lead everywhere. d1-3B scores higher on Fast Decisions (59.3), Typed Decisions (65.0) and AI2D (79.9, against 76.7 for d1-2B).

Intended use and limits

Liquid AI positions d1-2B for routing and triage, moderation and guardrails, intent and topic classification, extraction from documents and forms, judging and reranking, and visual inspection. It is not a chat model. Knowledge-heavy questions favor larger models, and answers get less reliable past about 64K tokens of state. Estimates are reliable on counts and scores, less so on free quantities. Liquid AI advises calibrating any decision threshold on data from your own use.

What this means

d1-2B replaces token-by-token generation with a single forward pass that outputs typed answers and probabilities. For pipeline steps such as classification, extraction and moderation, this means output formats are fixed by question type, and cost is not tied to generated length (usage reports zero output tokens). The one-call, many-questions design lets several decisions share one read of the same state. The benchmark claims come from Liquid AI's own board, and the model card itself says to calibrate thresholds on your own data.