最新模型
Perplexity Releases Decider v1.3 with Text Extraction and Character Positions
On November 12, Perplexity released Decider v1.3 and a Flash version, giving its decision models classification, scoring, and verbatim text extraction. The models use the Apache 2.0 license, but self-hosting requires a dedicated inference interface and validation of the effects of batch processing and probability calibration.

On November 12, Perplexity released Decider v1.3 and Decider Flash v1.3, adding Locate, a feature that extracts verbatim text from documents based on natural language instructions and returns the start and end character positions along with a score. For systems that need routing, content checks, or entity recognition, this offers another way to obtain structured results directly. Official announcement
Decider encodes the input and question, then reads option probabilities through a decision output layer, skipping the step of generating an answer token by token. In the new version, the full attention layers use bidirectional attention. Locate adds a span prediction layer and LoRA, enabled only during extraction; Perplexity found that merging them directly degraded the model’s existing decision-making ability. Technical explanation
The Locate interface guarantees that returned text corresponds to a slice of the original text, helping engineers link a decision back to document evidence. However, it currently accepts text only and cannot extract image content directly. Mixed classification and extraction questions still require separate forward passes, which add latency, and fast=True disables Locate. Choose the appropriate path based on the deployment workload. Model documentation
Performance figures also need to be interpreted according to how they were measured. On one million production requests, Perplexity used a large model to generate reference answers and reported 93.98% agreement for Decider v1.3, compared with 93.68% for the previous version. This measures agreement with a judge model, not accuracy against human annotations. Teams should still assess thresholds using their own Chinese documents and error costs. Evaluation methodology
For self-hosting, the main model’s BF16 weights require about 49 GiB of VRAM, while Flash requires about 8.5 GiB; both also need compute memory. Flash was distilled from the main model and shares its interface and features, making it suitable for initial testing of high-volume classification workflows. Main model card, Flash model card
Next, check compatibility with the serving implementation: the checkpoint uses a dedicated decision head, so default causal inference cannot reproduce its behavior, and the older prompt builder must be replaced. The official announcement says the weights are downloadable, but the model card still mentions a required access token. Confirm download access and complete local probability calibration before deployment. Deployment constraints