AI BriefingPerplexity Releases pplx-decider-v1-27b Decision Model Fine-Tuned from Qwen3.8-27B
Key point
The 26B parameter model achieves an 85.71% overall accuracy across 11 benchmarks, outperforming its base model Qwen3.8-27B and the Jev baseline.
Details
Perplexity has released pplx-decider-v1-27b, a decision model fine-tuned from Qwen3.8-27B. The model is designed for classification and decision-making tasks, providing calibrated probabilities for choices and yes/no queries.
Benchmark Performance
The model's results were measured through the Perplexity API across 11 benchmarks. It achieved an 85.71% overall accuracy, surpassing the base Qwen3.8-27B score of 74.76% and the Jev baseline of 84.51%.
Key performance metrics include:
- RAGTruth: 88.80% (vs 61.53% for base)
- FinancialPhraseBank: 84.18% (vs 75.68% for base)
- TabFact: 90.60% (vs 78.60% for base)
- WinoGrande: 83.30% (vs 73.10% for base)
Usage and Requirements
Running the model requires Python 3.12+ and a CUDA GPU with approximately 49 GiB of memory for weights and working memory. The model uses BF16 tensor types and has 26B parameters.
Users can download the inference example via uv and run it directly. The Decider class supports:
- Choice tasks: Selecting from defined criteria (e.g., routing support tickets).
- Yes/No tasks: Determining probabilities for binary questions.
- Image inputs: Passing images via the
predictmethod or command-line flags.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
GLM-5.3-Flash Matches Specialized Decision Models in Zero-Shot Classification via Logit Extraction
DoGBench: First User-Facing Documentation Benchmark Shows No Local Model Scores Above 50%