Perplexity Decider v1.3: Open-weight models that make decisions instead of writing text

What's going on

On November 12, 2026, AI search company Perplexity released two new models: pplx-decider-v1.3 and the smaller pplx-decider-flash-v1.3. Both ship under the permissive Apache 2.0 license and can be downloaded from Hugging Face, via the Hugging Face collection, the pplx-decider-v1.3 page and the pplx-decider-flash-v1.3 page. They're also available through OpenRouter and Perplexity's own Decisions API.

These aren't chatbots. Perplexity calls them "decision models." You give them some context, called the state, such as a customer email, a JSON document or an image, and ask one or more questions about it. What you get back isn't prose but a probability distribution over the possible answers. How likely is "yes"? Which of five teams should get this ticket? How upset does the customer sound on a scale from "calm" to "angry"? That makes them a good fit for the sorting and filtering jobs that come up constantly in software pipelines: routing requests, moderating content, reranking search results or grading the output of other AI models.

Two sizes, one idea

The large model is based on Qwen3.8-27B and needs about 49 GiB just for the weights in BF16, a common 16-bit number format. Flash is based on Qwen3.5-4B. The Reddit post's headline lists it as a 5B model. It needs about 8.5 GiB. Flash was created through distillation from the large model, meaning it was trained to imitate the answer distributions of the big "teacher" model. KL divergence measures how far apart the two distributions are. So Flash learns not only from correct answers but also from the teacher's uncertainty.

Under the hood, Perplexity replaces the usual text output, where the model predicts the next token one at a time, with a dedicated decision head. It's a small extra layer with 255 rows that gets read out at a marker position in the text. All answer options are scored in a single forward pass. That's why the model reports zero generated output tokens.

There's also an architectural change: the full-attention layers in v1.3 are bidirectional. When the model reads a word, it can also look at words that come later, not just earlier ones. The questions can't see each other directly, but the shared context can attend to the questions. In v1.2, an answer could still shift depending on what else you asked. The new approach is meant to make answers more stable. The catch for users: standard software such as plain Transformers loading or typical chat servers won't reproduce the model correctly. You need Perplexity's pplx_decider package.

The models were trained on hundreds of public datasets, converted into a unified format of context, question, options and target. According to Unite.ai, benchmark training splits were excluded. For later specialization on specific domains and on images, Perplexity uses LoRA, a method that keeps the base model frozen and trains only small add-on matrices. Perplexity credits much of the improvement to the "tasksource" dataset collection.

What the models can do

There are four question types:

Locate is new in v1.3. You describe what you're looking for, either as a short label like "person" or "email address" or as an instruction like "the person who signed the contract." The model then returns the exact text spans with character offsets, or an empty list if nothing matches. That's useful for named entity recognition (spotting names, places or companies), for finding personal data, and for extractive question answering, where the answer comes word for word from the text. Locate only works on text, not images, and you can ask at most 128 such questions per request.

Locate runs on its own LoRA adapter. Perplexity deliberately didn't merge it into the model because that would have hurt the regular decision tasks. The downside: if you mix Locate and decision questions, you need extra forward passes and wait longer.

The models handle images too. Each image is downscaled to at most about 1.6 megapixels and then takes up roughly 1,600 positions in the context. Training used sequences of up to about 30,720 tokens. The default limit is 32,768 tokens, and inputs are never truncated. Anything too long gets rejected.

Numbers and benchmarks

Most figures come from Perplexity's own tests:

nDCG@10 measures how well the top ten search results are ordered. According to Unite.ai, the large model also hits 91.6 percent mean accuracy across 63 query classification tasks, 6.54 percentage points above Perplexity's previous production classifiers. The reference labels for that test, however, were relabeled by a frontier language model. On the Artificial Analysis Index, both models sit on the Pareto frontier: no other model is both cheaper and better.

For a real-world test, Perplexity analyzed one million anonymized API requests from the API's first 18 days, containing 2,318,887 individual questions in total. Perplexity says users explicitly opted in. v1.3 agreed with the verdicts of large AI models 93.98 percent of the time, v1.2 93.68 percent. Unite.ai notes that's a small gap and only measures agreement with AI judgments, not accuracy against ground truth. Perplexity puts the cost of running this dataset at $26.47 (v1.3) and $13.24 (Flash). How local hardware was factored in isn't stated.

Internally, Perplexity uses Decider to review security logs from AI agents, among other things. On a single day, one billion tokens from 7,000 agent sessions ran through the model. That produced 14 suggested rule fixes, six of which were shipped within two days. Decider also filters low-quality web pages out of the crawl.

What you need to try it

The official route requires Python 3.12 or newer and an Nvidia GPU with CUDA. For the large model, the card needs room for about 49 GiB of weights plus working memory, which realistically means a 64 to 80 GB professional card or multiple GPUs. Flash gets by with about 8.5 GiB of weights, so it runs on a consumer card with 12 to 16 GB of VRAM. You'll need a similar amount of disk space. Macs and CPU-only setups aren't supported on the official route.

Perplexity doesn't list any quantized versions of v1.3. According to OrcaRouter, the Ollama library does have a model called "pplx-decider" as a 4-bit file of about 17 GB. But it's from an older version (v1 or v1.1). Which one exactly is unclear, and it has a much smaller context limit of 8,192 tokens. If you don't want to host v1.3 yourself, use OpenRouter or the Perplexity API.

The Model Card on Hugging Face still claims the repository is private. For older versions, OrcaRouter says that was an outdated notice.

The big picture

With Decider v1.3, Perplexity is going after a niche that gets little attention but is huge in practice: classifying, filtering and sorting, the countless small yes/no decisions inside AI pipelines. Instead of having a big chat model write text and then laboriously parsing it, you get probabilities directly in a single forward pass. That's faster, cheaper and easier to process. Locate also makes the model a competitor to tools like GLiNER that tag entities in text.

There are limits, though. Nearly all the numbers come from Perplexity itself, and the Artificial Analysis Index is one of the few external measurements. OrcaRouter tested the previous version, v1.1, on an independent leaderboard and showed where it falls short: it hit top scores on classic classification tasks but collapsed on hard multi-step reasoning (Humanity's Last Exam: 16.2). On top of that, about 2 percent of answers were wrong despite confidence above 95 percent. Perplexity itself advises calibrating thresholds on your own data. A single temperature value won't fit every task.

If you want to run the model yourself, you also have to use Perplexity's own code. Common tools like llama.cpp or vLLM won't reproduce it out of the box. For companies with large data volumes and their own GPUs, it's an attractive offer. For hobbyists, the Flash variant is the most interesting option.