Perplexity releases pplx-decider-v1.1-27b, outperforming Jev on Decision Index

Perplexity has released pplx-decider-v1.1-27b, an update to its decision-making model that improves the Decision Index score from 56.4 to 61.56. This performance now outperforms the Jev model by more than 3.5 points while utilizing the same Qwen3.8-27B backbone. The gains are primarily attributed to the lifting of causal attention constraints and the inclusion of additional training data from tasksource.

Evaluation

The overall score uses the suite's weighting, not a simple category average. Key category scores for pplx-decider-v1.1-27b compared to Jev and the previous version include:

• Knowledge: 48.18 (Jev: 51.4)
• Language: 69.45 (Jev: 62.0)
• Retrieval: 61.26 (Jev: 55.4)
• Tools: 78.88 (Jev: 75.1)
• Arts: 44.66 (Jev: 37.7)

Architecture and Serving

The checkpoint uses a native decision-checkpoint layout consisting of a Qwen3.8-27B backbone (identified as Qwen3_5Model in Transformers class name) plus readout.safetensors containing a BF16 [255, 5120] decision head.

Key serving requirements include:

• Full-attention layers use the saved noncausal attention mode. Default causal inference does not reproduce the evaluated setup.
• The supplied DecisionModel.predict applies the saved calibration temperature. When using logits directly, apply it exactly once.
• This artifact has a separate readout, not a full-vocabulary lm_head. Serving systems requiring Qwen3_5ForConditionalGeneration need a separate export mapping readout rows to vocabulary rows.

Usage

Usage requires Python 3.12+, authenticated Hugging Face access to the private repository, and a CUDA GPU with room for roughly 49 GiB of weights plus working memory.

Images

Images can be passed as file paths, PIL images, or base64 URLs. The processor resizes them preserving aspect ratio, rounding sides to multiples of 32 px.

• Token Limits: Text and images must fit within an 8,192-token limit.
• Resolution: Default max_image_pixels is 2048×2048. However, training and evaluation capped images at 512×512 (≤256 tokens). Higher resolutions fall outside the training distribution.