Decider v1.3 adds locate. The card still says private
Perplexity released Decider v1.3 and Flash v1.3 on November 12, 2026. They return probabilities and text spans. Hugging Face lists the weights as public.

Perplexity released Decider v1.3 and Flash v1.3 on November 12, 2026. They return probabilities and text spans. Hugging Face lists the weights as public.

On November 12, 2026, Perplexity published Perplexity Decider ↗. The post's citation date is that day. It releases Decider v1.3 and Decider Flash v1.3, successors to v1, v1.1, v1.2, and Flash v1.2. Flash v1.3 is a smaller model distilled from v1.3.
The picture above is ours. We have not run either checkpoint.
@perplexitydevs ↗ posted at 18:33 UTC: "v1.3 adds locate, a new feature that extracts spans based on type or instructions." The community announcement ↗ is timestamped 19:34 UTC. The API changelog ↗ files both models under November 2026.
A decision model, in Perplexity's definition, takes one input and a set of questions and returns a probability for each answer. It does not write a chat reply. v1.3 keeps three question types: yes/no (noul), multiple choice, and an ordered score. Locate is the new one. You name a field, or write an instruction such as "Extract the city name where the job is located," and the model returns the matching text plus start and end character positions. No match returns an empty list. One request can mix locate, classification, and scoring on the same input.
The pricing page ↗ lists pplx-decider-v1.3 at $0.02 per million input tokens and pplx-decider-flash-v1.3 at $0.01. Output tokens are free. There is no per-request fee. The aliases pplx-decider-latest and pplx-decider-flash-latest now select the v1.3 models and bill at those rates. v1.2 stays available if you name it.
Locate reads a string. The changelog says it does not work with web grounding. The v1.3 model card ↗ caps a locate request at 128 questions, with no images.
| Checkpoint | Published shape |
|---|---|
| v1.3 | Qwen3.8-27B. Card asks for about 49 GiB of BF16 weights, plus working memory |
| Flash v1.3 | Distilled onto Qwen3.5-4B. Card asks for about 8.5 GiB of BF16 weights |
| Decisions API | $0.02 and $0.01 per million input tokens. Output free |
| OpenRouter | Both :free routes listed at $0, 262,144-token context |
Both cards still say: "The repository is private: authenticate with a Hugging Face token that has access to it." On November 15 the Hugging Face model API listed both repos as public, not gated, under apache-2.0. The blog calls the weights downloadable under that license. The card's usage snippet has not caught up. We have not downloaded the files.
Local load, per the cards, wants Python 3.12 or newer and a CUDA GPU. Full-attention layers are noncausal. The card says default causal inference does not reproduce the checkpoint. The Hub tag on v1.3 is a finetune of Qwen/Qwen3.8-27B. Flash's tag is a finetune of Qwen/Qwen3.5-4B.
These figures are Perplexity's, from the November 12 post.
On 17 out-of-domain extraction tasks, Decider v1.3 scores 69.0 and Flash scores 63.6. Perplexity puts Liquid d1 at 43.3, Cloudflare's Clairic at 41.5, and GLiNER large v2.5 at 33.6.
On one million Decisions API requests from the API's first 18 days, v1.3 agrees with frontier-LLM judges 93.98% of the time, against 93.68% for v1.2. Those labels are the judges' answers. The requests had user permission for research.
Internally, Perplexity says Decider read 1 billion tokens across 7,000 agent sessions in one day and that the flaws it found led to 14 proposed rule fixes, 6 of them approved and shipped within two days.
"The latest versions of @perplexity_ai's open-weight decision models, Decider V1.3 and Decider Flash V1.3, are live on OpenRouter, and exclusively free for our users until November 19!" — Models on OpenRouter ↗ (@ModelsOnOR), November 12 at 19:54 UTC. The OpenRouter model page ↗ lists the price as free and does not print that end date.
"In one day it read 1B tokens across 7,000 agent sessions and found flaws that led to 14 fixes to their agent security rules. Both models are Apache-licensed and free on OpenRouter until Nov 19" — Alex Atallah ↗ (@alexatallah), November 13. The token count and the 14 fixes are Perplexity's, repeated here.
The @perplexitydevs post quoted above is the product announcement. Locate is the line that post leads with.
Reddit was not reachable for this story.
Use it when the code wants a probability, not a paragraph. Routing, moderation, rubric grading, and pulling a span out of a page are the jobs the blog describes. Perplexity points v1.3 at costly judgments and Flash at high-volume routing. The Decisions API price is the number to budget: $0.02 or $0.01 per million input tokens, output free. The OpenRouter :free routes are the trial, and @ModelsOnOR put the end of that trial at November 19. Today is November 15. Re-check the model page if you are reading this later. A chat SDK does not speak this API. OpenRouter's page says the free route uses its Decisions endpoint.
Read the card and the Hub listing together before you plan a download. Apache 2.0 and a public, ungated repo are what the API returned. "The repository is private" is what the card still says. Budget about 49 GiB for v1.3 or about 8.5 GiB for Flash, plus working memory, on a CUDA GPU. That is the card's floor, not a laptop claim, and we have not loaded the weights.
v1.2 thresholds do not carry over. The changelog says so. Locate is new, text-only, and off if you turn on web grounding. Gemini's Decisions API is the other shipping product in this shape: Decision 1, priced in that note at $0.10 per million input tokens, with yes/no, choice, and score, and no span extractor. Arete-27B is a different Qwen3.8-27B fine-tune. It still generates text. Decider replaces the language-model head with a decision head.
This is not a self-hosted agent. Anthropic's self-hosted Claude Managed Agents beta, opened the same day, runs a harness. Decider answers a question about a state. Chat plan comparisons are in ChatGPT vs Claude vs Gemini. The Qwen bases sit in the open-source LLM comparison.