The Pulse
Perplexity Releases Apache-Licensed 27B Decision Model
Perplexity’s open-weight pplx-decider-v1-27b is a Qwen3.8-27B fine-tune designed to return choices and probabilities for tasks such as classification and routing. Its model card reports an 85.71% overall score across 11 benchmarks, with Dec
AI.info Team · Written with AI ·
Perplexity released the open-weight pplx-decider-v1-27b on Hugging Face on October 1. Fine-tuned from Qwen3.8-27B, the model is intended to choose among predefined answers rather than compose chat responses. Its model card points to classification, routing and other tasks where software needs a choice or probability instead of open-ended prose. The listing carries an Apache 2.0 license.
A model built to choose, not chat
Perplexity, the model’s publisher, describes the release this way:
“pplx-decider-v1-27b is a decision model fine-tuned from Qwen3.8-27B.”
The model card’s examples show how it can be used for different decision formats. One asks which team should handle a customer’s Stripe integration problem, with billing, technical support and sales offered as choices. Another asks whether a message expresses urgency and returns a yes-or-no probability. For the multiple-choice example, the card says the prediction returns a selected choice and calibrated probabilities.
The model also accepts images. The usage instructions show an image file passed alongside a prediction request, including a screenshot example. Hugging Face labels the model as multimodal and lists it under text classification. These features let a developer provide visual material along with a question, while keeping the output tied to a defined task rather than asking for a general-purpose response.
Perplexity reports 85.71% across 11 tests
The model card reports an overall score of 85.71% across 11 benchmarks. Its comparison columns show 84.51% for Jev and 74.76% for the Qwen3.8-27B base model. Perplexity says Decider’s results were measured through its API; the card does not provide sample counts or a fuller evaluation protocol. The figures are therefore the company’s published results, not an independent comparison.
The row-level scores vary by task. Decider records 88.80% on RAGTruth and 90.60% on TabFact, ahead of both listed comparisons on those rows. It scores 83.30% on WinoGrande, below Jev’s 90.70% but above Qwen3.8-27B’s 73.10%. On FinancialPhraseBank, Decider reaches 84.18%, compared with Jev’s 76.98% and Qwen3.8-27B’s 75.68%.
On JudgeBench, Decider posts 78.29%, close to Jev’s 78.57% and above the base model’s 68.86%. For BBH, its 82.80% is below Jev’s 94.27% but above Qwen3.8-27B’s 72.80%. On JevBench public hard, Decider scores 70.30%; Jev records 73.27%, and Qwen3.8-27B scores 72.28%.
The ContractNLI row lists Qwen3.8-27B and pplx-decider-v1-27b tied at 80.78%, while Jev scores 77.45%. On Circa, Decider records 89.20%, compared with 87.00% for the base model and 84.60% for Jev. The remaining results are 94.00% on Belebele and 85.40% on TruthfulQA binary; Jev scores 95.00% and 92.00% on those tests, respectively, while Qwen3.8-27B records 93.20% and 82.80%.
Local use calls for a large CUDA GPU
The weights are available in Safetensors format under the Apache 2.0 license. Perplexity’s instructions call for Python 3.12 or later and a CUDA GPU with approximately 49 GiB available for weights, plus working memory. Hugging Face lists the model at 26 billion parameters and BF16 tensor type, despite the “27b” name and its Qwen3.8-27B base designation.
The repository provides an inference script and examples for loading the model, making predictions and passing image files. The example instructions use the Hugging Face download tool and then run the script locally. Hugging Face also says the model is not currently deployed through its Inference Providers.