frontier-infra/jebadiah-4b-v2

Jebadiah 4B v2: A Specialized Decision Model

Jebadiah 4B v2, developed by Frontier Infra, is a 4.5 billion parameter System One style decision model built upon the Qwen/Qwen3.5-4B base. Unlike traditional generative language models, Jebadiah is specifically designed to answer typed questions by providing a probability distribution over predefined option labels, rather than generating free-form text. This makes it highly suitable for structured decision-making and classification tasks.

Key Capabilities

• Typed Question Answering: Handles three distinct question types: choice (select one from N options), noul (yes/no statements with P(yes) output), and score (place a state on an ordered rubric).
• Probability-Based Output: Provides confidence scores for decisions, allowing for nuanced interpretation and integration into systems requiring probabilistic outcomes.
• System One Compatibility: Offers a TypeSafe-compatible /v1/systemone API endpoint, enabling seamless integration with existing Jev clients.
• Efficient Inference: Designed for a single forward pass per question, making it efficient for real-time decision applications.
• Standard Transformers Architecture: Can be run on various platforms, including CPU, CUDA GPUs, and Apple Silicon (via MLX builds).

Good For

• Automated Decision Systems: Ideal for applications requiring structured, probabilistic decisions, such as routing customer support tickets, classifying inputs, or scoring items based on criteria.
• Integration with Jev Clients: Developers already using Jev clients can easily switch to Jebadiah by changing the API endpoint.
• Resource-Constrained Environments: The 4.5B parameter size makes it suitable for deployment in environments where larger generative models might be too resource-intensive.
• Benchmarking and Evaluation: Useful for evaluating decision-making performance on specific datasets, with reported accuracy metrics on Jevals PubMedQA, Banking77, and Nimble evaluations.