locate is the addition that matters in pplx-decider-v1.3 and pplx-decider-flash-v1.3, Perplexity’s latest open-weight decision models. Ask for a type, or write an instruction, and the model hands back the matching stretches of the input, each tagged with character offsets and a score. The checkpoints are in the Decisions API and on Hugging Face under the Apache 2.0 licence. OpenRouter is running both for free until 19 November.
We’re rolling out pplx-decider-v1.3 and pplx-decider-flash-v1.3, the latest version of our open-weight decision models.
November 12, 2026
v1.3 adds locate, a new feature that extracts spans based on type or instructions.
Now available in our Decisions API, @huggingface, and @OpenRouter.
You hand them the content as state (text, JSON or images) and one or more named questions. What you get is a probability your code can act on: yes or no (noul), one of the options you defined (choice), or a level on an ordered rubric (score). Perplexity says it took its cue from Jev, TypeSafe’s System One model. The larger v1.3 checkpoint sits on Qwen3.8-27B. Flash v1.3, at 4B, is a distillation of that checkpoint onto Qwen3.5-4B.
Locate
Earlier checkpoints will reject a locate question; only the v1.3 models answer it. You can describe what to find with a short label such as “email address”, or with an instruction such as “Extract the city where the job is located.” Each hit comes back as a span: the verbatim substring, start and end offsets in Unicode characters (with end exclusive, so state[start:end] is the text), and a score. If the model finds nothing, you get an empty list.
In Perplexity’s documentation, a review that mentioned an AeroPress and the address [email protected] came back with those two spans at scores above 0.999 when asked for product names and email addresses. You can mix locate with noul, choice and score questions about the same state. Bo Wang, who wrote Perplexity’s evaluation post, said locate lets a decision model do named-entity recognition, pick out personal information, and answer extractive questions.
We introduced locate in decider v1.3. Now decision model can act as a NER, PII and perform extractive QA tasks.
November 12, 2026
We also wrote a blogpost demonstrating the training, evaluation and inspiring use cases of decider.
A request that includes a locate question has tighter limits. The state must be a string, with no images, at most 128 questions, and no web grounding. Training for locate used sequences of up to 8,000 tokens, and the published evals go as far as 32,000, Perplexity says. You can send longer inputs, though it hasn’t measured span quality past that point. On the self-hosted checkpoints, locate uses a separate adapter that stays unmerged, because merging it caused some regression on ordinary decision requests. Mixed requests therefore take separate forward passes for locate and for the other question types.
Prices, aliases and grouping
The Decisions API still bills the way v1.2 did: $0.02 per million input tokens for the 27B model and $0.01 for Flash. Output tokens are free, and there’s no per-request fee. Web grounding still adds $0.001 per search or successful URL fetch. The aliases pplx-decider-latest and pplx-decider-flash-latest now select the v1.3 models. Perplexity says probabilities can change with the model, so thresholds tuned on v1.2 should be revalidated. The v1.2 models remain available if you name them explicitly.
The way questions share a request has also changed. In the API, v1.3 and Flash v1.3 answer several questions about the same state in one pass, with the state billed once, but each question is isolated: it sees the state and its own text, not the other questions. On v1.2, questions could see one another, which could help or hurt depending on the grouping. Flash v1.2 still answers each question on its own and bills the state once per question.
Perplexity’s tests put a request of about 243,000 input tokens at 16 seconds on v1.3 and 5 seconds on Flash v1.3. The default organisation rate limit is 10 requests per second, shared across models, with 15 million tokens a minute on the 27B model and 4 million on Flash.
What Perplexity reports
Perplexity published a long evaluation post alongside the release. On one million deduplicated, PII-masked Decisions API requests from the API’s first 18 days, comprising 2,318,887 questions and included with user permission for research, it says pplx-decider-v1.3 had the highest agreement with frontier LLM judges of the models it tested, at 93.98%, compared with 93.68% for v1.2 and 93.35% for Jev 1.14. Production requests don’t come with ground-truth labels, so those figures measure agreement with the judges rather than accuracy against a gold set.
On internal production classification tasks, both models outperforms our existing classifier and the larger model outperforms Jev 1.14.
November 12, 2026
On one million real Decisions API requests, pplx-decider-v1.3 had the highest agreement with frontier LLM judges of any model test, at 93.98%.
On 63 internal query-classification tasks, with references independently relabelled by a frontier LLM, it reports 91.6% mean accuracy for v1.3, 6.54 percentage points above its production classifiers. On 17 out-of-domain extraction tasks used to test locate, it says v1.3 reached an average score of 69.0 and Flash v1.3 63.6, ahead of Liquid d1, Cloudflare’s Clairic and GLiNER large v2.5, using each task’s primary metric. It also reports a 0.27% answer-flip rate for v1.3 when labels were reordered or renamed on 372 multiple-choice questions, compared with 2.69% to 4.84% for Jev 1.14, and no changes when identical inputs were repeated.
Where to run them
You can also run the checkpoints yourself. They’re public in Hugging Face’s Decider v1.3 collection under Apache 2.0, with a Python 3.12+ loader that needs a CUDA GPU. The v1.3 card asks for about 49 GiB of BF16 weights plus working memory; Flash needs about 8.5 GiB.
OpenRouter has listed both as free until 19 November, under the ids perplexity/pplx-decider-v1.3:free and perplexity/pplx-decider-flash-v1.3:free, with a 262,144-token context window. locate runs on OpenRouter’s Extractions route. Free endpoints there are rate-limited, and OpenRouter’s documentation sets a lower daily cap unless you’ve bought credits.
pplx-decider-v1.3 and pplx-decider-flash-v1.3 are free on @OpenRouter until November 19.
November 12, 2026
How Perplexity is using it
Perplexity says it already uses Decider internally. In one security workflow, rule-based checks watch agent sessions in real time, and Decider reviews samples of the logs offline to find missed issues and false alarms. Those records go to an agent for triage, and a proposed rule fix is sent to a member of the security team for approval. In one day, Perplexity says Decider read 1 billion tokens across 7,000 agent sessions and identified flaws that led to 14 rule fixes, six of which were approved and shipped within two days. When those fixes were run back over a week of logs, traffic on the busiest rule fell by 52%, and the widened rules picked up about 40% more legitimate actions that were worth a look. Adel Ka, who leads detection and response at Perplexity, said the team is using it to tighten Numbat, the company’s endpoint agent monitoring rules. Kyle Polley, the chief information security officer, wrote that this open-source decision model already watches those rules in real time and helps improve them.
✨ pplx-decider-v1.3, the latest version of our open decision model, is out!
November 12, 2026
great blog on how @perplexity_ai trains, evaluates, and uses it internally, including a security use case where we use it to improve our Numbat (endpoint agent monitoring) rules.
we’re already using it for other security use cases too. excited to share more soon!
It also uses Decider to monitor reinforcement-learning rollouts, answering questions about failed tool calls, code that ran without errors, and final answers, and storing probabilities rather than a verdict so thresholds can be changed later. Mark Tenenholtz, a member of technical staff, said the team consistently sees “better and more reliable results than using smaller API models like Luna or Haiku, and it’s *way* cheaper.” The blog says the monitor only observes and diagnoses: it never changes a reward or stops a run.
We use Decider internally to monitor our RL rollouts.
November 12, 2026
We consistently see better and more reliable results than using smaller API models like Luna or Haiku, and it’s *way* cheaper.
A third internal use is page-quality filtering as pages are crawled for the index, with other classifications such as domain, audience, and whether a page contains harmful, illegal or personally identifiable information run in the same pass. The official documentation still recommends v1.3 for nuanced judgments, detailed rubrics, and workflows where errors are costly, and Flash v1.3 for high-volume routing, labelling and filtering. It also tells users to tune thresholds on a labelled sample of their own traffic, rather than assume a fixed cutoff will hold across tasks.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
