Decision Models

Decision models are a new class of AI model purpose-built for structured decisions. Instead of generating text token by token, a decision model evaluates a situation and returns calibrated probabilities across a fixed set of outcomes in a single call with zero generated tokens.

d1 Models   Latest release

Which d1 Should I Use?

Pricing

The hosted models are billed on input tokens only. Output is free because decision models do not generate tokens. Each question is billed as its own prompt, including its text and all images. See Image Token Usage for how images are counted. The models are available through the Liquid Console. d1-ultra and the earlier d1 model are also on OpenRouter and Vercel AI Gateway. See the Liquid AI pricing page for current prices and available credits. d1-2B, d1-omni-2B, and d1-omni-600M are open-weight models under the LFM Open License and free to run on your own hardware.

Primitives

Every question you ask a decision model uses one of three core types:

• Noul: A yes/no question that returns a probability between 0 and 1. “Is this message spam?” might return 0.92.
• Choice: A pick-one-from-many question that returns a probability distribution over named options.
• Score: A rate-on-a-scale question that returns a probability-weighted position on an ordered rubric.

Typed Answers

d1-2B and d1-omni-2B add typed answers on top of the three core types. Every question has a type and instructions, and every answer repeats its type.

• Missing values: a value or a box is missing when found is under 0.5. Its value, text, and span are then None. For boxes, instances under 0.5 are left out.
• Values are read, not converted: number and date are parsed from the span. For example, 4 % reads as 4 with the unit %. A value that is found but can’t be parsed keeps its span, and the value is None.
• JSON states: when the state is JSON, a found value also has a path, the RFC 6901 pointer to it.
• Estimates: the model reads an estimate on at most 128 values between low and high. Estimates have no found. They are reliable on counts and scores, less so on free quantities.
• at_least: True when every instance the model read was present, so the picture may hold more than count.
• Boxes need one picture: a request with a box or boxes question must have exactly one picture in media.
• Runtimes: typed answers run in Transformers and mlx-vlm. llama.cpp supports Noul, Choice, and Score only.

Calibrate any decision threshold on data from your own use.

When to Use Decision Models

Decision models are a good fit when the answer is a structured decision: classification, categorization, routing (tickets, emails, requests), scoring, triage, content moderation, guardrails and safety checks, LLM-as-judge replacement, agent tool-call approval, and model routing or cascades. With image input, the same primitives support visual inspection, image classification, and decisions based on screenshots. With audio (d1-omni-2B and d1-omni-600M), they support voice-command routing and spoken-intent detection. Use a language model instead for free-form text generation, creative writing, multi-turn conversation, complex multi-step reasoning, open-ended Q&A, summarization, and code generation. To move existing LLM calls to a decision model, see the Decision Model Guide.

Examples