Skip to content
NLEN
Illustration: a small glowing chip on a circuit board sorts streams of documents into three coloured lanes with probability bars, with a large server block in the background.

Perplexity Decider v1.3: open 27B and 4B decision models, now with locate

By Ivo Donker · — text: Grok Bot (AI), checked against the sources by Claude (AI)

What is Perplexity Decider v1.3? It is a pair of decision models Perplexity released on November 12, 2026: pplx-decider-v1.3 with 27 billion parameters and the smaller pplx-decider-flash-v1.3 with 4 billion. They do not write text. They return probabilities for questions about a piece of input, such as yes or no, a pick from options or a score on a scale. New in v1.3 is locate, which pulls verbatim text spans out of the input. The weights are on Hugging Face under an Apache 2.0 license.

We read the model cards on Hugging Face, Perplexity's pricing page and changelog, and the model page on OpenRouter. The topic was also in the trending briefing of November 13.

What Perplexity released

Perplexity announced the models on November 12 on its community forum. According to the pplx-decider-v1.3 model card, the large model is built on Qwen3.8-27B. The Flash model card says it runs on Qwen3.5-4B and is distilled from the 27B model. Both accept text and images.

A request holds a state (the input) and one or more questions. There are four question types. noul returns the probability of yes. choice returns a probability per option, with up to 255 options. score returns an expected position on an ordered scale. The new locate returns text spans with start and end offsets and a score. Locate works on text only, without images, and with at most 128 questions per request. The model produces no output tokens.

Run it yourself or use an API

For local use, the model card asks for a CUDA GPU with room for about 49 GiB of BF16 weights for the 27B model and about 8.5 GiB for Flash, plus working memory. The cards still say the repository is private, but on November 13 the Hugging Face API listed both models as public and not gated, with an apache-2.0 license. We covered shrinking models to fit less memory earlier in our piece on quantization in production.

In Perplexity's Decisions API, v1.3 costs 0.02 dollars per million input tokens and Flash 0.01 dollars, according to the pricing page. Output tokens are free and there is no per-request fee. The aliases pplx-decider-latest and pplx-decider-flash-latest now point to v1.3. On OpenRouter, Flash v1.3 is listed as a free variant with a 262,144-token context window. That variant uses OpenRouter's Decisions API; a regular chat SDK does not work with it. Perplexity said on X that both models are free on OpenRouter until November 19. OpenRouter's model page shows a price of zero but gives no end date.

What this means

A decision model fits in front of a more expensive language model in an agent: it can route, filter or label requests at a fraction of the price, and the application sets its own threshold on the returned probability. The 4B version is small enough for a single GPU with about 9 GiB free. It fits the wider move to open weights we track in our overview of open-source LLM trends. Anyone who tuned thresholds on v1.2 should test them again, because an alias moves to the new model automatically.

What we did not check

We did not run the models ourselves. We could not open Perplexity's blog post with benchmarks and training details, so we leave out accuracy figures. The November 19 end date for free use on OpenRouter comes only from a Perplexity post on X.

Checked on November 13, 2026.

This article was written with AI (Grok Bot) and checked against the sources below.

Sources