Perplexity has released Decider v1.3 and Decider Flash v1.3, two classification models that return probabilities instead of generated text, and has put the weights on Hugging Face under an Apache license. Both are also reachable through OpenRouter and the Perplexity Decisions API. Developers can wire those probabilities into sorting, routing, ranking, and grading steps.
Flash is the smaller sibling, distilled from the larger model. The relationship to other labs is specific. Perplexity says Decider is inspired by Jev, the System One Model from TypeSafe AI, which means it is a separate model and not a fine-tune of Jev. The backbones are Qwen models: the larger Decider is based on Qwen3.8-27B, and Flash is trained from it into the smaller Qwen3.5-4B.
The new capability is called locate. A user describes what to find, either with a short label such as “location” or with an instruction, and the model returns the matching text from the input along with start and end character positions. A miss yields an empty list. Perplexity trained it on named entity recognition, personal information detection, and extractive question answering. A single request can mix locate questions with classification and scoring about the same input.
Locate comes with a trade-off that Perplexity states openly. Folding the locate adapters into the base weights degraded ordinary decision requests, so the company leaves them unmerged and activates them only for locate calls. Teams that self-host should expect that extra moving part.
On extraction, Perplexity tested 17 out-of-domain tasks and compared the models with three extractors: GLiNER large v2.5, Liquid d1, and Cloudflare’s Clairic. Decider v1.3 averaged 69.0 and Flash averaged 63.6. Liquid d1 scored 43.3, Clairic 41.5, and GLiNER 33.6. By Perplexity’s account, neither model lost a single category.
The comparison against Jev, the nearest rival in design, is tighter. In a sample of one million production requests to the Decision API, Decider v1.3 agreed with frontier LLM judges 93.98 percent of the time, against 93.35 percent for Jev 1.14. Jev won relevance ranking by 0.69 points. Decider won by 0.93 points where it counts most by volume, information extraction, and by 1.21 on intent and routing.
Retrieval results are modest. Across 72 retrieval tasks, Decider v1.3 reached a mean nDCG@10 of 83.99, ahead of Qwen3-Reranker-8B by 1.81 points and Jev 1.14 by 2.24. Flash trails in finance, programming, and law when set against Jev, so the smaller model is not a free swap everywhere.
Stability is the sharper claim. Perplexity shuffled options and renamed labels across 372 multiple-choice questions. Decider v1.3 flipped its answer on 0.27 percent of them under each transformation, while Jev 1.14 flipped on 2.69 to 4.84 percent. Repeating identical inputs produced no changes from Decider and a 1.88 percent change rate from Jev. These are Perplexity’s own tests, and the company concedes that public leaderboards can be gamed, which is why it built an internal suite of more than 200 tasks.
Perplexity also describes how it uses the model. Inside Perplexity, a single day of Decider reviewing agent logs covered 7,000 sessions and 1 billion tokens, and it surfaced flaws that produced 14 security rule fixes. Six shipped within two days. The company says the same forward pass can also filter low-quality crawled pages, and it recommends Flash for high-volume routing and v1.3 where errors are costly.
The cost chart in the post carries its own warning: the figures depend on the workload, local hardware costs are not itemized, and latency was not compared like for like. Anyone replacing an LLM judge with Decider should price their own traffic and tune thresholds on a small labeled set, as Perplexity advises, before trusting its published margins.
Based on Perplexity’s blog post “Perplexity Decider v1.3: Training, eval, and use cases,” which carries no publication date.
