Perplexity releases pplx-decider-v1.2 decision model with joint request format and higher accuracy

Perplexity has released pplx-decider-v1.2, a new version of its decision model built on the Qwen3.8-27B backbone that scores 68.09 on the overall Decision Index, up from 62.30 for the previous version. The jump comes from two changes: a joint request format that lets multiple questions share context in one pass, and substantially reworked training data that includes official benchmark training splits for fairer comparisons.

The joint request format is the headline change. Instead of answering one question per forward pass, v1.2 reads the state and every question in a single sequence. Each question gets its own readout at a marker token placed after its options. The result: answering several questions about the same state costs roughly the same as answering one, and the questions can inform each other. On SATA-Bench, where users select all applicable options, grouping questions lifted scores by 11.3 points. On aspect-category sentiment analysis, the gain was 7.8 points.

But grouping is not always better. When questions are independent candidates competing for the same judgment - like selecting one function call from several candidates - the joint format hurts. On function calling benchmarks, scores dropped by 1.1 points. On reasoning-intensive retrieval, they fell by 2.9 points. The model's own documentation is blunt about the trade-off: "Grouping helps when the questions are facets of one decision, and hurts when they are independent candidates competing for the same judgement."

How the training data changed

The team reworked the training data substantially. Several benchmarks in the Decision Index ship official training splits, and models that train on them gain markedly on those benchmarks. Perplexity now includes these splits, decontaminated against the test sets, to make comparisons fair. The company said it "did not optimize for that alone" and added a large amount of other decision data while tracking every benchmark for regressions.

The checkpoint handles both text and images up to 1.6 megapixels. On the Vision Decision Index, it scored 77.43, up from 73.22 for v1.1. The final model combines several training runs through model souping - weight averaging chosen as the best compromise across all evaluations rather than any single benchmark.

When to group questions and when to split them

The model's behavior on grouped versus separate questions is well-documented. For related questions - several attributes of the same item, multi-label decisions, or routing steps - one request is both more accurate and much cheaper than sending N separate ones. For independent candidates, the recommendation is to score them in separate requests or small chunks if cost matters more than the last point of accuracy.

On BEIR reranking with 13 datasets, the macro effect of grouping 8 candidate documents per request was a slight drop of 0.5 points. But the range was wide: multi-hop datasets gained 2.1 points, while counter-argument datasets lost 2.8 points. The guidance: "When in doubt, use one candidate per request: it is never more than about 0.4 points worse, while grouping can cost about 3."

Technical details and throughput

The model requires Python 3.12 and a CUDA GPU with about 52 GiB of BF16 weights plus working memory. A fast path using torch.compile and CUDA graphs pushes throughput to 21.2 requests per second on Decision Index workloads, a 2.33× improvement over the reference path. With FP8 matmuls enabled, that rises to 24.9 requests per second, though probabilities drift slightly from the reference - argmax agreement drops from 99.8% to 99.4%.

Temperature is fitted at 1.382 on held-out calibration data. On the development set of 1,497 questions, the calibrated model achieves 91.3% accuracy with an expected calibration error of 0.014. On held-out decision datasets, probabilities remain somewhat over-confident, with ECE around 0.05 - better than v1.1 but not perfectly calibrated.

The model uses noncausal attention over the full request, with the readout applied at marker tokens rather than the last token. Default causal inference will not reproduce the checkpoint's behavior. The repository is private and requires a Hugging Face token with access.

Why this matters for IT and development teams

For teams building classification, routing, or moderation pipelines, the joint request format changes the cost calculus. A request with 10 questions costs about the same as one question, which means multi-label decisions that previously required 10 API calls can now run in one. But the accuracy trade-offs are real and task-specific. The model's documentation provides clear heuristics for when to group and when to split - and the fast path's packing and compilation options mean separate requests stay cheap even at scale. Before adopting, test both modes on your own data, particularly if your task involves ranking independent candidates.