Introduction - Perplexity releases new decision models with Locate feature
Perplexity has launched new versions of its open-weight decision models, Decider v1.3 and Flash v1.3. These models are designed to improve how AI systems make decisions based on text and images. The main update is a new feature called Locate, which extracts exact text spans from inputs. This release aims to enhance accuracy and usability for tasks like named entity recognition (NER), personal identifiable information (PII) detection, and question answering (QA). These models are available on Hugging Face, the Perplexity Decisions API, and OpenRouter. They are built to run efficiently on modern hardware and support various deployment options.
The models are part of Perplexity’s family of decision models, which focus on returning probabilities rather than generating text. This approach makes them suitable for classification, scoring, and reranking tasks. The new Locate feature allows users to pinpoint specific parts of a text, making the models more versatile. The models are open-weight, meaning their weights are freely available under the Apache 2.0 license. This openness supports research and custom deployment. The models are designed to handle large inputs and deliver fast responses, making them useful for real-world applications.
Model details - Sizes, architecture, and training methods
Decider v1.3 is a decision model with 27 billion parameters, based on a large language model (LLM) backbone. It is trained using a full fine-tuning process on hundreds of public datasets. The training involves averaging weights across three different seeds to improve stability. The model’s architecture includes full-attention layers that are made bidirectional. This means the model can consider context from both directions in a sequence, which improves understanding.
The Flash v1.3 model is smaller, with 4 billion parameters, and is distilled from the larger model. It is trained by minimizing the Kullback-Leibler divergence, a method that helps it mimic the larger model’s outputs. Both models are designed to run on hardware with a CUDA GPU that has around 49 GiB of memory for v1.3 or about 8.5 GiB for Flash. There are no quantized variants of these models, so they require substantial hardware resources. The models are optimized for high performance and accuracy in decision tasks.
The models’ training process involves pruning the language model’s head to focus on option tokens like A, B, and C. A linear decision head reads the hidden state at a special marker token and produces probabilities for each option. This setup allows the models to answer yes/no questions, multiple-choice questions, and score questions effectively. The attention mechanisms are adjusted so that each question only sees the state and itself, not all other questions, which helps maintain consistency across multiple questions in a single request.
Performance benchmarks - Accuracy, domain, multilingual, and robustness
Perplexity reports that Decider v1.3 performs well across a variety of tasks. It achieves a mean accuracy of 91.6% on 63 query classification tasks. This score is 6.54 points higher than their previous models used in production. The model also leads in domain classification across seven different areas, with improvements of up to 12.9 points in finance. It excels in retrieval and reranking tasks, with a mean nDCG@10 score of nearly 84, which surpasses other models like Qwen3-Reranker-8B and Jev 1.14.
The models also perform strongly in multilingual tasks, with an accuracy of 88.16% across 19 tasks in eight languages. For out-of-domain tasks, Decider v1.3 averages 69.0 in Locate, which is higher than Flash’s 63.6. In image classification, the model scores nearly 77%, outperforming Jev 1.14. On real-world requests, the models show high agreement with human judges, with nearly 94% accuracy in a test of over 2 million questions.
Robustness is another key feature. The models have a low answer flip rate of 0.27%, meaning they rarely change answers when inputs are reordered or relabeled. This is significantly better than Jev 1.14, which flips answers at a rate between 2.69% and 4.84%. These benchmarks are self-reported by Perplexity, so actual performance may vary in different settings.
Locate feature - Exact span extraction for NER, PII, and QA
The Locate feature is a major addition in Decider v1.3. It allows the models to extract exact text spans from inputs, which is useful for named entity recognition, PII detection, and extractive question answering. Users can specify fields with labels like "location" or detailed instructions such as "Extract the city name where the job is located." The models then return the start and end character offsets for each span.
To support this, a new special token called <|locate|> is added for each requested field. The models are trained with a span head and a Low-Rank Adaptation (LoRA) adapter, while keeping the backbone and decision head frozen. The adapter stays unmerged because merging it reduces decision accuracy. This setup allows the models to precisely identify parts of a text without retraining the entire model.
The Locate feature enhances the models’ ability to handle tasks that require exact extraction. It is particularly useful in scenarios where understanding specific entities or data points is critical. The training process fine-tunes the span head to identify start and end positions accurately, improving the model’s usefulness in real-world applications.
Deployment options - API, local use, and hardware requirements
Decider v1.3 and Flash v1.3 are available through multiple deployment options. Users can access the models via the Perplexity Decisions API, which supports high-volume requests and offers a straightforward integration. The API endpoint is a POST request to a specific URL, with pricing set at $0.02 per million tokens for v1.3 and $0.01 for Flash.
For local deployment, the models can be loaded using the Decider.from_pretrained() method from the pplx_decider package. Running these models locally requires a CUDA GPU with sufficient memory. For v1.3, about 49 GiB of memory is recommended, while Flash needs around 8.5 GiB. Both models require non-causal full-attention layers to function correctly. No quantized versions are available, so hardware must be capable of handling the full models.
These deployment options give flexibility for different use cases. The API is suitable for scalable, cloud-based applications, while local use allows for more control and privacy. The hardware requirements are high but manageable with modern GPUs. The models are designed to deliver fast responses, with latency times reported at around 16 seconds for large requests on v1.3 and 5 seconds on Flash.
Comparison with other models - Jev 1.14, Qwen3-Reranker-8B
Perplexity compares Decider v1.3 with other decision models like Jev 1.14 and Qwen3-Reranker-8B. In self-reported benchmarks, v1.3 scores 91.6% on query classification, outperforming Jev 1.14 by a significant margin. It also leads in domain classification across seven domains, especially in finance, where it scores higher than competitors.
In retrieval and reranking tasks, v1.3 achieves a mean nDCG@10 of nearly 84, which is better than both Jev 1.14 and Qwen3-Reranker-8B. For multilingual tasks, v1.3 maintains a high accuracy of over 88%, surpassing Jev 1.14. In image classification, it scores nearly 77%, slightly above Jev 1.14. The models’ robustness also favors v1.3, with a lower answer flip rate.
Latency tests show v1.3 takes about 16 seconds for a large request of approximately 243,000 tokens, while Flash completes requests in about 5 seconds. Pricing is competitive, with v1.3 costing $0.02 per million tokens and Flash costing $0.01. These comparisons highlight v1.3’s strengths in accuracy, domain coverage, and speed.
Pricing and latency - Costs, speed, and request limits
The API pricing for Decider v1.3 is set at $0.02 per million input tokens, while Flash v1.3 costs $0.01 per million tokens. Output responses are free of charge. This pricing structure makes the models accessible for large-scale applications. The models are optimized for speed, with latency times around 16 seconds for large requests on v1.3 and about 5 seconds on Flash.
Request limits are designed to support high-volume use cases. The API can handle requests with up to 256,000 tokens of input. For the Locate feature, requests can include up to 128 questions, enabling complex extraction tasks. The hardware requirements are high, with v1.3 needing about 49 GiB of GPU memory and Flash around 8.5 GiB. These specifications ensure that the models can deliver responses quickly and reliably.
Cost and speed are balanced to support both real-time and batch processing. The models’ performance and pricing make them suitable for deployment in production environments where efficiency and accuracy are critical.
Why it matters - Impact on classification, retrieval, and safety
These models improve decision-making in AI systems. Their high accuracy in classification tasks helps in sorting and understanding large amounts of data. The Locate feature allows precise extraction of information, which is useful for compliance, data privacy, and information retrieval.
The models also enhance retrieval and reranking, making search results more relevant. Their robustness reduces errors caused by input reordering or relabeling, increasing safety and reliability. This is important in sensitive applications like finance, law, and healthcare, where mistakes can have serious consequences.
Overall, Decider v1.3 and Flash v1.3 support safer, more accurate AI systems. They enable better handling of complex tasks and large data volumes. Their open weights and flexible deployment options make them accessible for research and custom solutions. This update helps AI systems become more precise and trustworthy in real-world use cases.
How OpenSmartRoute helps
Routing requests through OpenSmartRoute allows teams to optimize how they use these models. The platform can direct each request to the best-fit model, agent, or tool based on quality, speed, and cost. It can pin requests to on-premises models for sensitive data, ensuring safety and compliance.
OpenSmartRoute learns from outcomes, so models that perform well get more traffic. It can also set rules to stay within region or cost limits. The platform supports open-weight models like Decider v1.3 and Flash v1.3, making it easier to deploy and manage them at scale. It provides detailed cost tracking and quality scoring, helping teams optimize their AI workflows.
By integrating these models into its routing system, OpenSmartRoute helps teams reduce costs, improve response quality, and ensure safety. It simplifies managing multiple models and deployment options, making AI more accessible and reliable.
Source: https://www.marktechpost.com/2026/11/13/perplexity-decider-v1-3-open-weight-decision-models-locate/