frontier-infra/jebadiah-27b-v2-1

Jebadiah 27B v2.1: A Specialized Decision Model

Jebadiah 27B v2.1, developed by frontier-infra, is a 27 billion parameter model built upon the Qwen/Qwen3.8-27B base. Unlike traditional chat models, Jebadiah is specifically engineered as a decision model. It processes typed questions and outputs probabilities over predefined option labels (choice, noul for P(yes), and score based on a rubric) in a single forward pass, without generating conversational text.

Key Capabilities

• Probabilistic Decision Output: Provides direct probabilities for given options, ideal for classification and structured decision tasks.
• Optimized for Label Logits: Reads label-token logits efficiently with "thinking off" for fast, direct answers.
• Performance: Achieves a Decision Index 0.3 public-only score of 57.03, showing improvements across Knowledge & Reasoning (+2.34), Language Understanding (+3.89), Retrieval & Classification (+2.81), and Arts & Human Taste (+1.31) compared to its predecessor.
• T1e Inference Recipe: When run with the T1e recipe (repeating payload, reversing options for low-confidence answers), the public-only index rises to 60.46, though this increases latency.
• Low Latency: Median in-process request latency of 106.5 ms on an RTX PRO 6000 for standard reads.

Good for

• Applications requiring direct, probabilistic answers to multiple-choice or scoring questions.
• Classification tasks where output needs to be a probability distribution over labels.
• Systems needing to interpret user input into structured decisions rather than free-form text generation.
• Use cases where a smaller, faster 9B version is insufficient and higher decision accuracy is paramount, provided memory resources are available.