Perplexity pplx-decider v1.3: Locate Spans and Flash Distill

Perplexity on November 12, 2026 shipped Decider v1.3 and Flash v1.3—successors to the October pplx-decider launch with a locate extraction route, refreshed training, and Apache weights on Hugging Face. Decision models return probabilities, not chat prose, so routers can threshold yes/no, multi-class, rubric scores, and now text spans in one HTTP contract.

The release sits next to Google’s Decision 1 API and explainx.ai’s decision-model roundup as structured routing becomes a first-class product category.

TL;DR

What a decision model is in Perplexity’s framing

Perplexity positions Decider as inspired by Jev and TypeSafe’s System One Model—a model that reads state (text, JSON, or images) and answers named questions with typed probability objects:

Application code—not the model—picks thresholds. That pattern avoids brittle JSON parsing from chat completions and matches how Interfaze extraction models target pipelines, but Decider keeps everything inside one Decisions API surface.

Locate: when spans beat generative extraction

Generative models often paraphrase extracted values. Locate is Perplexity’s bid for verbatim spans with coordinates—useful for:

• Named entity recognition with audit trails
• PII detection where compliance wants exact offsets
• Extractive QA where downstream systems highlight UI text

When no match exists, Perplexity says locate returns an empty list rather than hallucinating a span—a design choice aligned with high-precision compliance workflows.

Structured decision outputs let routers threshold probabilities instead of parsing chat text

Full v1.3 versus Flash v1.3

Perplexity recommends v1.3 when errors are costly and rubrics are nuanced, and Flash v1.3 for high-volume routing where latency and dollars dominate.

Secondary write-ups of the hub post describe:

Flash is distilled from the teacher using KL divergence on output probabilities, not only hard labels—standard teacher-student practice but worth noting when you fine-tune further.

Evaluation claims—and what to rerun yourself

The hub blog reports 91.6% mean accuracy across 63 query-classification tasks with references relabeled by a frontier LLM judge. On one million sampled production API requests (2,318,887 questions), Perplexity reports 93.98% agreement with frontier LLM judges for Decider v1.3.

Those numbers are vendor-reported. Before you replace production classifiers:

1. Sample your label distribution.
2. Tune thresholds on a hold-out set (Perplexity explicitly recommends this).
3. Compare latency and cost against Flash and against hosted Jev baselines you already run.

On 17 out-of-domain extraction tasks, Perplexity claims both v1.3 models beat general extraction models—again, rerun on your documents.

How to try v1.3 this week

Hosted (fastest smoke test).

• OpenRouter exposes perplexity/pplx-decider-v1.3:free with posted zero token pricing November 12 and sub-second latency in the provider table.
• Perplexity’s Decisions API remains the first-party path if you already integrated October’s $0.04/M input pricing.

Self-host.

• Weights: perplexity-ai/pplx-decider-v1.3 and pplx-decider-flash-v1.3 on Hugging Face.
• Expect CUDA and RAM beyond raw parameter counts; Flash is the realistic on-prem starting point.

Consult the hub post’s Getting started section for request shapes for locate versus legacy decision types.

What people are asking

Training and architecture changes in v1.3

Perplexity’s hub post attributes gains to bidirectional attention, better-curated decision training data, and a refined training recipe versus v1.2. Bidirectional context matters when state includes long policies or multi-paragraph terms-of-service: a unidirectional decoder may underuse left context when scoring a clause at the end of a document.

Flash v1.3 inherits those teacher logits through distillation, trading some nuance for sub-second OpenRouter latencies (~0.38s in the provider table we saw November 12). For agent routers that fire dozens of micro-decisions per user turn, Flash is the economic default; keep full v1.3 for appeal review, safety escalation, and finance approvals.

Sample decision workflows

Pair Decider with general LLMs the way Perplexity describes internally: cheap decision passes filter or route before you spend frontier tokens on synthesis.

Cost planning: OpenRouter free tier to production

OpenRouter’s free v1.3 endpoint is useful for labeling experiments, not indefinite production—trackers cite a November 19, 2026 end. When you migrate to paid routing:

• Compare Perplexity Decisions API invoice rates against October’s $0.04/M input baseline.
• Model Flash for always-on classifiers and full v1.3 for nightly batch relabel jobs.
• Cache identical state hashes if your platform allows—decision models are deterministic at a given temperature setting.

Self-hosters should budget GPU hours for 26B only if latency requirements beat API economics; Flash at ~5B parameters is the usual middle ground.

Relationship to Jev and open decision-model ecosystem

Perplexity still credits Jev / System One inspiration. explainx.ai’s October guide showed pplx-decider-v1-27b slightly ahead of Jev on an 11-benchmark panel but behind on hardest rows. v1.3’s 91.6% claim is a new panel—not comparable by headline number alone.

If you standardize on one decision schema across vendors, wrap Perplexity, Google Decision 1, and open weights behind an internal router interface so benchmarks can swap without rewriting application code.

Migration notes from v1.2

Teams on pplx-decider-flash-v1.2 should:

1. Re-tune thresholds—probability calibrations shift between versions.
2. Test locate on redacted logs before enabling on production PII.
3. Update Hugging Face revision pins in Dockerfiles to the v1.3 tags.
4. Run shadow traffic on OpenRouter free tier before cutting over paid traffic.

Limits

• Locate quality on noisy OCR or scanned PDFs is not broken out in the hub post we read.
• Independent reproduction of 91.6% panel scores requires the task list and judge prompts Perplexity did not fully export in the announcement.
• Context at 262K on OpenRouter does not imply your self-hosted deployment matches KV memory without profiling.

Related on explainx.ai

• Original pplx-decider v1 and Decisions API
• Top decision models ranked
• Google Decision 1 API
• Interfaze extraction model
• Structured output patterns
• Perplexity embed v2 preview

Locate API expectations for engineers

Perplexity routes locate through an Extractions path distinct from probability-only decisions. Expect payloads that name a field or instruction, return character offsets into the original state string, and attach confidence per span. Empty lists mean no match, not “model refused.”

That contract suits highlight UI in reviewer tools and audit logs in regulated industries better than free-form JSON blobs from chat models. It also plays nicely with structured output guides when you compose Decider output into larger agent graphs.

Hardware notes for self-hosting

Secondary summaries cite ~49 GiB BF16 for full v1.3 and ~8.5 GiB for Flash—before KV cache and framework overhead. A single consumer GPU may serve Flash for batch offline labeling; full v1.3 typically wants datacenter GPUs at modest batch sizes. Profile with your longest state strings because 262K context on OpenRouter marketing does not guarantee your hardware configuration matches.

When not to use Decider

Skip Decider when you need open-ended summarization, creative writing, or multi-step tool planning. Use it when the output space is small and scored—routing, moderation labels, compliance checks, extraction spans. Misapplied decision models waste tokens because they never articulate reasoning chains unless you bolt a separate explainer model on top.

Monitoring after you cut over

Log per-route latency, threshold rejection rates, and locate empty-list frequency separately. A spike in empty locate results often means instruction drift or encoding mismatches in upstream state, not model regression. Alert when disagreement rates exceed your baseline. Keep a v1.2 shadow for one week if you already had production traffic—compare disagreements carefully before decommissioning older weights.

Citation block for papers and internal docs

Perplexity published a BibTeX entry on the hub post dated 2026-11-12 with authors spanning Perplexity and ML contributors. Cite that entry when you reproduce locate benchmarks internally so version v1.3 is explicit—not the October v1 weights.

Release details follow Perplexity’s hub blog and Hugging Face model cards as of November 12, 2026.