Not tested by us yet. Decider v1.3 is a weights release with no hosted endpoint, and the model card says no inference provider serves it. Everything below comes from the pplx-decider-v1.3 model card on Hugging Face. We'll update this post when the weights are openly downloadable or a hosted route appears, so we can run it against other decision models on our own items.
Perplexity Decider v1.3 is a decision model: it does not write text, it scores answers to typed questions about an input. It is a 26B-parameter finetune of Qwen3.8-27B under Apache-2.0, and its new feature is Locate, which returns verbatim text spans with character offsets. If you want a self-hosted decision model that can also point at the evidence, it is the first one we have seen that does both. If you want a hosted API with a price per token, look at Gemini Decision 1 instead.
| Decider v1.3 | |
|---|---|
| Maker | Perplexity AI |
| Registered by us | 2026-11-13 (the card shows no release date) |
| Parameters | 26B, BF16 |
| Base model | Qwen/Qwen3.8-27B |
| License | Apache-2.0 |
| Question types | choice, noul (yes/no), score, locate |
| Inputs | Text or JSON state, plus optional images |
| Hardware | CUDA GPU with about 49 GiB for the BF16 weights plus working memory, Python 3.12+ |
| Access | Hugging Face repository that needs a token with access |
| Price | No hosted API, so no per-token price |
What a decision model is
A chat model answers by generating tokens that you then have to parse. A decision model skips that step. You send a state and a set of questions, and it returns probabilities. The card lists the four question types:
- choice: one option out of 1 to 255 labeled options, with per-option probabilities and a confidence.
- noul: a yes/no question, returned as P(true).
- score: a position on an ordered scale of 2 to 255 levels, with the expected level and probabilities.
- locate: matching text spans pulled from the state, each with
start,end,textandscore.
The first three match the question types in Kanana-2-v-decide and Salesforce's Koa Decisions. A request has a state (a string, dict or list), a questions object keyed by ID, and optional images. The response is one answer object per question ID, plus usage.
What is new
Locate is the headline. The earlier decision models we have covered answer "which one" or "how much". Locate answers "where", which suits jobs like finding the clause that triggered a policy flag or the sentence that supports a routing choice. The card describes it as an extraction capability and gives no accuracy figure for it.
The card calls v1.3 the next version of pplx-decider-v1.2 with improved performance and reliability across decision tasks. It gives no numbers behind that line and points readers to a blog post for evaluation, which we have not been able to use as a source here.
There is also a speed note. The card has two request modes. In separate mode (the default) each question runs in its own sequence. In isolated mode the whole request is encoded as one sequence, which is cheaper when you ask many questions but changes 1.5% to 2.2% of top answers compared with separate. Perplexity's fast path is reported as about 1.7 to 1.9x faster than default padded batches in separate mode (99.6% to 99.7% top-answer agreement), and 9 to 18x faster than the reference path in isolated mode (99.4% to 99.8% agreement). Those are Perplexity's figures on its own validation sample, and they measure agreement with its reference code, not accuracy on your data.
How it compares on paper
| Decider v1.3 | Strands Decider 2B | TypeSafe Jev | |
|---|---|---|---|
| Size | 26B | About 2B on Qwen3.5-2B | Not stated |
| Weights | Yes, Apache-2.0, gated repo | Yes, on Hugging Face | No, hosted early access |
| Price | Your GPU | Your hardware | $0.042 per 1M input tokens, output free |
| Latency claim | None on the card | Under 150 ms on a local machine | 70 to 500 ms end to end |
| Extraction | Locate spans | Not described | Extraction listed among uses |
The Strands and Jev figures come from AWS's launch coverage and TypeSafe's announcement. They are vendor claims, and none of the three vendors publishes a comparison against the others that we could verify. The size gap is the practical story: a 2B model fits on a laptop, while 49 GiB of weights means a data-center GPU.
Who should care
- Teams that cannot send data to a hosted API. Apache-2.0 weights let you run decisions inside your own network, if you can get access to the repository.
- Anyone building evidence-linked review flows. Locate returns offsets, so a reviewer can see why a flag fired.
- Not for latency-sensitive routing on small hardware. A 26B model is the wrong tool for a per-request routing step when a 2B decision model or a hosted one exists.
The access detail is the main catch. The card says you need a Hugging Face token with access to the repository, which sits oddly beside an Apache-2.0 label, and we could not confirm that anyone outside Perplexity can download it today. Treat it as announced, not yet generally available, until a clean download works.
What we will test when we can run it
One set of 30 to 60 items, with control items that have obvious answers and a mirrored wording for each yes/no question, because a model that flips when the question is reversed cannot be trusted for routing. We will put it next to Jev, Strands Decider 2B if it runs, and a cheap and a mid-priced chat model acting as classifiers, and report accuracy, controls passed, median latency and cost per 1,000 decisions. For where decision models sit among recent releases, start with the Gemini Decision 1 write-up.