[KongchangAI]
· 2 min read· 1,284 words

Perplexity Releases Decider v1.3: Decision Model Adds Locate Capability

Perplexity Releases Decider v1.3: Decision Model Adds Locate Capability

Perplexity releases open-weight decision models Decider/Flash v1.3, adding span extraction capability.

Perplexity has released Decider v1.3 and Flash v1.3, two open-weight decision models that output probability distributions and structured determination results instead of natural language, designed for routing, classification, and span localization in retrieval-augmented pipelines. v1.3 adds Locate for span extraction, sharing input representations with classification to reduce model switches. It reports 91.6% average accuracy across 63 query classification tasks but lacks per-category breakdowns and separate Locate evaluations. Open weights enable local deployment and fine-tuning, though key metrics like parameter scale and inference cost remain undisclosed.

Perplexity has released Decider v1.3 and Flash v1.3, two open-weight decision models positioned differently from mainstream chat models. Their core difference lies in the output target: instead of generating natural language, they directly return probability distributions and structured determination results. Version v1.3 adds the Locate capability, which can be used for span extraction, and reports an average accuracy of 91.6% across 63 query classification tasks.

The Division of Labor Between Decision Models and Generative Models

Making "decision-making" a separate model layer is an inevitable outcome of the evolution of retrieval-augmented products like Perplexity. Chat models excel at organizing and expressing information, but in intermediate steps such as routing, classification, and span localization, generating text token by token is both slow and introduces unnecessary randomness. Decider and Flash output probabilities rather than text, essentially handing a determination problem to a model form better suited to it: given an input, return category confidence and target positions.

The direct benefit of this design is pipeline controllability. Upper-layer applications receive comparable numerical values rather than natural language answers that need to be parsed again, and errors are easier to localize to a specific step rather than being buried in an entire generated passage. Open weights allow teams to fine-tune on their own data or deploy locally, which is friendlier for scenarios sensitive to latency and data boundaries.

Locate: Turning "Find It" into a Native Model Capability

The most noteworthy addition in v1.3 is Locate. Span extraction is extremely common in retrieval and question-answering pipelines: locating the original sentence in a long document that supports a conclusion, or identifying key entities in a user query. In the past, such tasks were often handled by a separate extraction model, or by adding a layer of rule-based post-processing on top of a generative model.

Locate incorporates this step into the decision model itself, meaning classification determination and span localization can share the same input representation. For pipelines that need to simultaneously answer "what intent does this query belong to" and "where is the key span corresponding to that intent," reducing one model switch can lower overall latency and integration complexity. As for extraction quality, the official release currently only provides overall accuracy on classification tasks; separate evaluation data for Locate has not yet been disclosed, and actual performance still needs to be verified on specific corpora.

Span extraction is a fundamental task in information extraction: given a piece of text and a question or intent, the model must return a contiguous character range (i.e., a span) in the original text as the answer, rather than regenerating a passage. Typical applications include answer localization in reading comprehension, entity recognition, and evidence sentence extraction in retrieval systems. Compared with generative answers, span extraction has the advantage that answers come strictly from the original text, are traceable, less prone to hallucination, and have a smaller output space with faster inference. Traditional approaches usually adopt extractive models based on encoders such as BERT, predicting probability distributions over start and end positions on the input. Locate builds this capability into the decision model, meaning classification and extraction can share the same encoded representation, reducing model switches in the pipeline.

What 91.6% Accuracy Means

The 91.6% average accuracy corresponds to 63 query classification tasks. The informational value of this number depends on the composition of the task set: average accuracy reflects the overall level but masks variance across tasks. If task difficulty is unevenly distributed, weaknesses in a few high-difficulty categories may be diluted by the mean.

A more pragmatic reading is to treat it as a baseline reference. The value of decision models lies not only in point accuracy but also in the adjustability brought by probability outputs: low-confidence samples can be escalated to humans or stronger models for review, thereby achieving higher effective accuracy at the system level. The engineering value of such "classifiers with confidence" in routing scenarios is often more critical than absolute precision numbers.

Average accuracy is the arithmetic mean of accuracy across multiple tasks, which differs from "micro-average" or "macro-average": if sample sizes vary greatly across tasks, a simple average may overestimate or underestimate overall performance. Reporting a single mean across 63 tasks leaves readers unable to judge whether every task is close to 91.6%, or whether some tasks are near perfect while others are significantly lower. When evaluating classification models, one typically also needs to look at confusion matrices, per-category precision/recall/F1, and confidence calibration—that is, whether the probabilities the model provides truly reflect its correctness rate. For decision models, calibration quality directly determines the effectiveness of a "low-confidence escalation to human" strategy, and is therefore more valuable for engineering reference than a single accuracy figure.

Deployment Choices Brought by Open Weights

Open weights are another main thread of these two models. They allow the models to be downloaded, audited, and modified, making them suitable for teams that require data to stay within their domain or have strict constraints on inference costs. Compared with closed APIs, the trade-off is that teams must bear deployment, version management, and evaluation work themselves.

The dual-model combination of Decider and Flash also suggests a layered approach: one handles heavier determination tasks, while the other targets lightweight scenarios. However, official materials do not disclose key metrics such as parameter scale, context length, and inference cost, so selection still requires hands-on testing against one's own workload.

Open-weight means model parameters are publicly downloadable, but training data, training code, or the full training process may not be open, so it is not equivalent to a fully open-source model. Users can deploy locally or in private clouds, audit weights, perform fine-tuning, or quantization and compression, thereby meeting requirements for data staying within their domain, low latency, and controllable costs. The trade-off is bearing the burden of inference infrastructure, version upgrades, security assessment, and continuous evaluation. In contrast, the closed API model has the service provider handle operations, but data must leave the domain, and behavior may drift as the provider updates. Selection requires weighing compliance requirements, team operations capability, and total cost of ownership.

Applicable Scenarios and Areas to Watch

In the short term, such models are better suited to being embedded in intermediate layers such as retrieval, routing, content moderation, and structured information extraction, rather than directly facing end users. For teams already building RAG or Agent pipelines, the decision model layer can replace part of prompt engineering and rule-based code.

Two things are worth watching going forward: the stability of Locate on long documents and multilingual corpora, and the per-category breakdown of performance across the 63-task set. If the official team fills in these evaluations, Decider's positioning will shift from "an interesting model form" to a component that can be seriously considered for architecture selection.

One-line Summary

Perplexity releases open-weight decision models Decider/Flash v1.3, adding span extraction capability.

Paragraph Summary

Perplexity has released Decider v1.3 and Flash v1.3, two open-weight decision models that do not generate natural language but instead directly output probability distributions and structured determination results, designed specifically for intermediate steps such as routing, classification, and span localization in retrieval-augmented pipelines. Version v1.3 adds the Locate capability, supporting span extraction and sharing input representations with classification determination, reducing model switches in the pipeline. The official release reports 91.6% average accuracy across 63 query classification tasks, but lacks per-category breakdown data and separate evaluation of Locate. Open weights facilitate local deployment and fine-tuning, suitable for scenarios sensitive to data boundaries, but key metrics such as parameter scale, context length, and inference cost have not yet been disclosed. Overall, such models are better suited to being embedded in RAG or Agent pipelines, replacing part of prompt engineering and rule-based code.

#Perplexity#Decider v1.3#Flash v1.3#open-weight model#span extraction#decision model#RAG pipeline
Share:

Related articles