Decision models, explained

A decision model reads an input and a set of questions and returns a probability for each answer in one pass, without writing text. Use one when you need a pick, a yes/no or a score, not an essay. Below, four decision models compared using only what each model card says.

When to use a decision model

Use a decision model for one-pass picks: choose one option from a named list, like which team gets a support ticket.
Use it for yes/no checks: is this a refund request, does this image contain a chart, does this reply follow policy.
Use it for scores: rate urgency, relevance or quality on a fixed scale, with a probability for each level.
Use it for routing: send each request to the right model, tool or queue before any text is generated.
Use it for screening: filter content, flag risky inputs or rerank search results at high volume.
Use a chat model instead when you need written output: drafts, explanations, summaries or multi-step reasoning you want to read.
Check the license and hardware needs on the model card, then test on your own labeled examples before trusting any score.

Compared: what each model card says

From each model card	JEV-35B ↗	d3 (Decision 3.0) ↗	d1-3B ↗	pplx-decider-v1.1-27b ↗
Made by	AutoTrust	vLLM Semantic Router team	Liquid AI	Perplexity
License	Apache-2.0	Apache-2.0	Other (LFM 1.0 license): not open-source	Apache-2.0
Size	35B mixture-of-experts, about 3B active per token	26.09B; the family spans 0.59B to 26.09B in six models	3.12B	27B (Qwen3.8-27B base)
Inputs	Text and images	Text or JSON, images and video	Text, JSON and images	Text and images
Answer types	Choice, yes/no and score; up to 256 options in one pass	Choice, yes/no and score, several questions per call	Yes/no, choice and score (2 to 10 levels), zero output tokens	Choice and yes/no decisions
Speed (author's own claim)	Median 204 ms per request on one B200	Median 54.4 ms per text request on one AMD MI325X	8 ms per decision on an RTX 4090	Not stated
Hardware (per card)	About 66 GiB GPU memory in bf16; NVFP4 build about 23 GiB	About 62 to 65 GiB GPU memory per replica	Card reports speeds on an Apple M5 Pro and NVIDIA Jetson boards	About 49 GiB of weights plus working memory; card says access needs authentication
Score (author's own claim)	Decision Index 0.3 public suite 59.73, own scoring, not a board entry	Jev Decision Index 0.3.2: 66.3, ranked first on the text board	Decision Index 0.2.1: 48.57, own scoring, not a leaderboard submission	Decision Index 61.56, up from 56.4 for v1

Every score and speed figure above is the model author's own claim from its model card, measured on different hardware and different Decision Index versions, so the numbers are not directly comparable. None has been independently replicated by AgentPlays.

Prompts

AgentPlays template Written by AgentPlays because the source doesn't publish an exact prompt we could quote. Fill in the [brackets].

{"team": {"type": "choice", "instructions": "Which team should handle this [ticket]?", "criteria": {"billing": "[Charges, refunds, invoices]", "technical": "[Setup and faults]", "sales": "[Buying questions]"}}, "urgent": {"type": "noul", "instructions": "Is the customer blocked right now?"}, "quality": {"type": "score", "instructions": "How complete is the request?", "criteria": ["[Missing key details]", "[Some details]", "[Everything needed]"]}}

Watch out: Decision models answer only the questions you give them and can be confidently wrong. Benchmark scores are each author's own claim. Check every model's license on its card before commercial use: d1-3B uses Liquid AI's LFM 1.0 license, which is not open-source.