A 4B Perplexity decision model that returns calibrated probabilities for routing, labeling and filtering instead of generated text.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Eight Points Behind the 27B Sibling, Built for Single-Question Gates
Technical Specifications
What the model supports
Comparison
pplx-decider-flash-v1.2 or pplx-decider-v1.2: When Is 4B Enough?
These two models share the same Perplexity Decisions API contract, text and image state, and Apache 2.0 weights, so switching is a one-line model change. They differ in size (about 4B versus 27B), per-token price and multi-question behavior. pplx-decider-v1.2 answers several questions in one pass and bills the state once. Flash answers each question separately and bills the state again each time. The decision depends on how many questions you ask per input and how costly a wrong answer is.
Choose pplx-decider-flash-v1.2 when you ask one question per input, the criteria are clear (routing, labeling, filtering), volume is high, and the lower per-token rate or a small self-hosted footprint matters. Choose pplx-decider-v1.2 when you grade against detailed rubrics, check claims against long source text, or face costly errors. Also choose it when you ask three or more questions about the same state, where Perplexity says it usually costs less. Perplexity's own comparison shows the 27B model ahead on both text and vision.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Ticket and message routing
Flash returns a probability for each of up to 255 options plus the most likely one, which suits routing messages to teams, queues or intents. Perplexity names routing and labeling as its core use for Flash and offers a support-ticket triage cookbook. Application code thresholds the result instead of parsing text. Send one question per request. For multi-facet triage on the same message, the 27B sibling bills the state once and handles grouped questions better.
Pre-filter gate before costly models
Place Flash upstream of generation models to filter spam, moderate content, or decide whether an input needs a full agent. Because output is a probability, you can auto-act at high confidence and escalate the uncertain middle to pplx-decider-v1.2, a larger model or a human. The 4B size and input-only billing suit high-volume traffic. Calibrate thresholds on labeled data, since the benchmark gap versus the 27B model is real.
Document and image classification
Flash takes images as well as text and JSON, resized to at most 1.6 megapixels, with up to eight images per question. That fits tasks such as identifying document types or checking whether required fields are filled. Perplexity reports 72.53 on its vision suite, versus 77.43 for the 27B model. Fine-grained visual judgments or safety-critical image checks should be validated carefully or moved to the larger model.
Relevance scoring and reranking
Perplexity says v1.2 added reranking capability, and the model card advises scoring each candidate (document, tool, call) in its own request, which is Flash's native mode. Perplexity's reported retrieval category shows the smallest gap to the 27B model (64.8 versus 68.4, joint runs). Use it to judge whether a passage answers a query before generation. Check ranking quality on your own corpus first.