Dek: A decision model doesn't write you an answer. It hands back a probability for each option it was given, which is why five questions about one support ticket cost one pass instead of five.
What decision models do that a chatbot can't
A decision model is a classifier wearing an LLM's clothes. You give it a document and a set of labelled options, and it returns a probability for each option rather than a paragraph of prose.
The payoff is cheaper than it sounds. One forward pass answers every question you attach to the same document, so a routing layer can ask whether a ticket is billing, whether it's urgent, and whether it needs a person, without three separate calls to a chat model.
Narrow on purpose.
That's the pitch, and it's a smaller tool than a chatbot. Nobody is going to hold a conversation with this thing, and anyone who tries will conclude it's broken.
Perplexity Decider v1.3 ships Apache 2.0 weights
Perplexity published the release post on 12 November for pplx-decider-v1.3 and pplx-decider-flash-v1.3, successors to v1.2 and Flash v1.2. Weights sit on Hugging Face under Apache 2.0, and you can call the models through the Perplexity Decisions API or through OpenRouter.
The large model is adapted from Qwen3.8-27B. Flash v1.3 is a 4B model distilled from it on top of Qwen3.5-4B. Perplexity's post says 27B and 4B, the Hugging Face listing shows 26B and 5B, and the gap is worth knowing if you're counting memory before you download.
Supported questions are yes or no, multiple choice, and scores. That's the fine print people skip, and it explains most of the disappointment when someone points a decision model at an open-ended task and gets a number they can't use.
The attention change that keeps Decider v1.3 answers stable
v1.3 changes how questions see each other. The shared state sees all of them. Each question sees only that state and itself.
In plain terms, questions share the document but stop watching each other, so adding a fourth question no longer nudges the answer to the first one. Perplexity reports a 0.27 percent answer-flip rate across label and format perturbations. That means the same document with a reworded question lands on the same label almost every time.
That number matters more than an accuracy point. An unstable classifier is worse than a mediocre one, because you can't tell which of your million rows to trust.
Locate turns the open-weight classifier into a span finder
Locate is new in v1.3, and it does a different job. A special token, a span head and a LoRA adapter let the model pull spans out of a document and return character positions, for named entity recognition, personal data detection and extractive question answering. Locate requests accept up to 128 questions.
The adapters ship unmerged, and Perplexity says why: merging them caused a regression on ordinary decision requests. That's an unusually candid note in a release post, and it saves a self-hoster from finding out the hard way.
The benchmark numbers are Perplexity's own
Perplexity reports 91.6 percent mean accuracy across 63 query-classification tasks. It says that's 6.54 points above its production classifiers, and that v1.3 leads TypeSafe's Jev 1.14 on all seven domain-classification sets, by as much as 12.9 points in finance.
On retrieval and reranking, the report gives 83.99 mean nDCG@10 across 72 tasks, a metric that rewards putting the right document near the top of a list.
Every one of those figures comes from the company that built the model. Nobody has published a reproduction. Read them as a vendor's homework, not a scoreboard.
How much of it survives contact with your own tickets? Nobody knows yet, and that's the only benchmark you'll care about.
The same tables carry a caveat: Flash trails Jev 1.14 in finance, in programming, and in law. Locate averages 69.0, the lowest absolute score in the post, which fits a feature that shipped this week.
Decider pricing, GPU memory and the self-hosting bill
The API costs $0.02 per million input tokens for v1.3 and $0.01 for Flash. Output is free, because there isn't any. Perplexity puts production requests at $26.47 per million, and that's the figure to take to whoever signs off on infrastructure.
Self-hosting is the other path, with a hardware price attached. v1.3 needs about 49 GiB of BF16 weights, Flash needs 8.5 GiB, and Perplexity published no quantised builds. The big model wants a serious GPU. The small one fits on a modest card, which is probably the more interesting option for a team testing an idea.
Context limits are 256K tokens per API request at most, with 32,768 as the local default.
Where Decider fits in a real pipeline
Routing, moderation and reranking are the obvious slots, and judge work fits the same shape. Anything where you'd otherwise ask a chat model the same question a million times and throw away the prose.
Perplexity uses the family internally for security rule mining, reinforcement-learning rollout monitoring and page-quality filtering. In one day it read 1 billion tokens across 7,000 agent sessions and surfaced 14 rule fixes, six of which shipped within two days. That's the vendor describing its own results, and it's still the most concrete example in the post.
The family is inspired by TypeSafe's Jev. The same week brought Google's Gemini Decision 1, a separate launch that Latent Space grouped into what it called decision models day. Two releases in seven days is a category forming, or two vendors agreeing on a name.
If you're building a routing layer, the free weights are the part worth your afternoon: benchmark them on your own tickets before you pay for anything. If you want an assistant that explains itself in sentences, this isn't the tool, and no amount of tuning will make it one. Decider classifies, scores and extracts. Judge it on that.






