Perplexity’s Decider v1.3 adds text extraction and downloadable model weights

Perplexity has released Decider v1.3 and Decider Flash v1.3, adding exact text-span extraction and downloadable Apache-2.0 weights. The update gives developers more ways to use the models for classification and retrieval without asking them to generate prose.

Perplexity Watch analysis

What happened

The new Locate feature returns matching text with its start and end character positions. Perplexity describes it for named-entity recognition, personal-information detection and extractive question answering; if there is no match, it returns an empty list. Locate is text-only and capped at 128 questions per request.

Decider produces probabilities for yes-or-no questions, custom choices or ordered scores. The company says v1.3 changes how questions attend to one another, aiming to prevent an answer changing simply because another question was included in the same request. Both models’ weights are downloadable under Apache 2.0, though self-hosting requires a CUDA GPU and memory beyond the model weights.

The larger Decider model has roughly 49 GiB of BF16 weights; Flash has roughly 8.5 GiB. Perplexity reports 91.6% mean accuracy for the larger model across 63 query-classification tasks, and 83.99 nDCG@10 across 72 retrieval tasks. For a sample of API requests, it reports 93.98% agreement with answers supplied by LLM judges. That last figure measures agreement with model-generated references, not accuracy against ground-truth labels. These details come from Unite.AI’s report on the release.

Why it matters

Decider targets work where an application needs a score or a chosen category, not a fluent paragraph: routing, moderation, filtering and reranking, for example. Locate adds a useful way to point to the evidence behind a decision, rather than merely returning a label.

There are practical trade-offs. Requests mixing Locate with decision questions require separate forward passes, which adds latency. And a model’s probabilities still need thresholds tuned to the application: the cost of a false positive in one workflow may be quite different from the cost of a false negative in another.

Our read

This is a substantive release for teams building classification or retrieval systems, with a welcome emphasis on extractable evidence and weights developers can run themselves. Start with Flash if hardware or throughput is the constraint, then test thresholds against labelled examples from the actual task. A benchmark is a useful starting point, not a substitute for your own awkward data.

What to watch

• Whether developers find Locate useful enough to justify the extra computation in mixed requests.
• How the models perform on application-specific, labelled data.
• Whether independent evaluations support Perplexity’s reported benchmark results.

Discussion spark: For moderation or routing, would you trust a model that returns probabilities and quoted evidence, or do you need independently verified accuracy before using it in production?

Independent WittyWires tracker for public updates about Perplexity. Not affiliated with or endorsed by Perplexity; this is not an official account.