Perplexity has released two new models in its Decider family, built for one job: classifying an input and choosing among predefined options. They do not generate text word by word. They receive a state, the content to examine, plus one or more questions, and return a probability distribution over the possible answers. The versions are v1.3, with about 26 billion parameters, and Flash v1.3, with about 5 billion, a smaller model distilled from the larger one.
How they work
The large model is based on Qwen3.8-27B; the compact version comes from Qwen3.5-4B and is trained to mimic the probability distributions of its bigger sibling. Both accept text and images. Running v1.3 locally needs about 49 GiB of GPU memory for the BF16 weights alone; Flash needs about 8.5 GiB. The models handle yes/no, multiple-choice and ordered-scale questions, and each answer comes back as a probability that an application can use to route, rank or filter.
The new feature: Locate
The new capability is Locate. Given a description, either a field type such as "person" or "city" or a natural-language instruction, the model pulls the matching passages out of the text, copied verbatim, with start and end character positions and a confidence score. The stated use cases are named-entity recognition, personal-data detection and question answering over text. Locate does not work on inputs containing images and accepts up to 128 questions per request.
Pricing and access
The weights are published on Hugging Face under an Apache 2.0 license, but the repository currently appears private and requires an access token, so it is not the open download the announcement implies. The models also run through Perplexity's Decisions API and on OpenRouter, where they are free for a week. In the API, the standard price is $0.02 per million input tokens for v1.3 and $0.01 for Flash, with output free.
Perplexity also offers guidance: the decision threshold should be tuned on a small labelled set rather than fixed in advance. v1.3 is recommended for nuanced judgements and workflows where errors are costly; Flash for high-volume routing and filtering where speed and cost matter more. Testing on one's own workload is still necessary, since training used sequences up to 30,720 tokens and inputs are not truncated.
The internal figure
In its announcement post the company reports two numbers from its own workloads: on 63 production classification tasks, v1.3 would be 6.54 points more accurate than the classifiers it currently runs; on a set of 17 external extraction tasks, the family would significantly outperform the best general model. The more practical point is the ability to replace a multi-component pipeline with a single pass. In the agent-security context, the same company says analysing one billion tokens across 7,000 sessions led to 14 rule fixes.
For developers building AI agents, the open question is whether a model that returns choices instead of prose really cuts latency and cost in real workflows. Perplexity says so; testing on one's own work is the way to check.

No comments yet. Be the first!