Perplexity open-sources Decider v1.3 with text-location outputs

• Open weights ship in 27B and 4B sizes under Apache 2.0, according to the release overview.
• Locate adds verbatim text extraction with character offsets, enabling batch decisions plus evidence extraction, as described in the quoted launch announcement.
• Mean accuracy reaches 91.6% across 63 internal classification tasks, 6.54 percentage points above Perplexity’s deployed classifiers, according to the launch post.

Exact offsets are the nicest detail in the API contract: they count Unicode characters, with an exclusive end, so Python’s state[start:end] reproduces the returned text. Ordinary causal inference cannot reproduce this checkpoint, warns the model card.

A Locate question describes what to extract, using a label such as “email address” or an instruction such as “Extract the city where the job is located.” Answers contain a spans list with start, end, text, and score; no match returns [].

The API documentation specifies these boundaries:

• Plain text only: Locate requires a string state. Images, objects, and arrays return 400, although ordinary decision questions support multimodal inputs.
• 128 questions maximum: extraction can share a request with binary, multiple-choice, and scoring questions.
• No web grounding: combining Locate with grounding returns 400.
• Limited long-context evaluation: Locate was trained on inputs up to 8K tokens and evaluated up to 32K. Longer inputs are accepted, but extraction quality beyond that range is unmeasured.

The production baseline is an embedding model feeding separate binary and multi-label classification heads, according to Perplexity’s evaluation report. References for the 63-task comparison were independently relabeled by a frontier LLM; the company describes its internal evaluations as zero-shot, without task-specific fine-tuning.

For Locate, Perplexity reports these averages across 17 out-of-domain extraction tasks:

| Model | Average score |
| --- | ---: |
| Decider v1.3 | 69.0 |
| Decider Flash v1.3 | 63.6 |
| Liquid d1 | 43.3 |
| Cloudflare Clairic | 41.5 |
| GLiNER large v2.5 | 33.6 |

Those averages combine different metrics: exact-span F1 for named entities and PII, character-level F1 for extractive QA, and field accuracy for structured extraction. They are not a single extraction-accuracy percentage.

Locate adds a span head and LoRA adapters while freezing the original backbone and decision head, Perplexity explains in its training write-up. Merging the adapters caused regressions on decision requests, so they remain unmerged and activate only for extraction.

Local inference has several consequential switches, documented in the checkpoint’s serving instructions:

• mode="separate", the default: each decision question gets its own sequence, including another read of the state. Reported evaluations used this mode.
• mode="isolated": questions share computation over the state. Because the state also attends to the questions, changing the question set can change answers; 1.5–2.2% of top answers differed from separate mode in Perplexity’s samples.
• Mixed extraction and decisions: one request still requires separate forward passes for the two question types, adding latency.
• fast=True: the optimized local path disables Locate.

pplx-decider-v1.3 uses Qwen3.8-27B; pplx-decider-flash-v1.3 is distilled onto Qwen3.5-4B.

The Decisions API documentation lists $0.02 per million input tokens for Decider v1.3 and $0.01 for Flash v1.3, with output tokens free. Separately, OpenRouter access is free for one week, according to the launch announcement.

An internal audit processed 1 billion tokens across 7,000 agent sessions in one day, producing 14 proposed security-rule fixes, according to Perplexity’s deployment account. Six were approved and shipped within two days.

Rule-based checks remain on the real-time path with negligible added latency; Decider reviews logs offline for missed issues and false alarms. Internal security applications are also feeding a data flywheel for further improvements, the team said in a follow-up.