frontier-infra/jebadiah-4b-v2

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Jebadiah 4B v2 by Frontier Infra is a 4.5 billion parameter open System One style decision model, built by merging a LoRA adapter into Qwen/Qwen3.5-4B. It specializes in answering typed questions (choice, yes/no, score) by providing a probability distribution over option labels rather than generating text, making it highly efficient for structured decision-making tasks. This model is optimized for integration into existing Jev client systems via a TypeSafe-compatible /v1/systemone API, offering a distinct approach to AI-driven decision support.

Loading preview...

Jebadiah 4B v2: A Specialized Decision Model

Jebadiah 4B v2, developed by Frontier Infra, is a 4.5 billion parameter open System One style decision model. Unlike traditional generative LLMs, Jebadiah is designed to answer typed questions by providing a probability distribution over predefined option labels, rather than generating free-form text. This makes it particularly well-suited for structured decision-making and classification tasks.

This model is built upon the Qwen/Qwen3.5-4B chat checkpoint, with a LoRA adapter merged into its bf16 weights. It processes questions in a single forward pass, returning precise probabilities for 'choice' (pick one of N), 'noul' (yes/no statements as P(yes)), and 'score' (placing a state on an ordered rubric) question types. It is compatible with existing Jev client systems through a TypeSafe-compatible /v1/systemone API endpoint.

Key Capabilities

  • Structured Decision-Making: Provides probabilistic answers for choice, yes/no, and scoring questions.
  • Efficient Inference: Answers questions in a single forward pass, without text generation.
  • API Compatibility: Integrates seamlessly with Jev client systems via a /v1/systemone endpoint.
  • Standard Transformers Model: Can be run on various platforms, including CPU, CUDA GPUs, and Apple Silicon.
  • Calibrated Outputs: Applies per-type temperature scaling for calibrated probability distributions.

Good For

  • Automated Triage and Routing: Categorizing customer inquiries or support tickets.
  • Probabilistic Classification: Applications requiring confidence scores for discrete choices.
  • System One Style Decision Support: Integrating into systems that require fast, structured judgments.
  • Resource-Constrained Environments: Its 4.5B parameter size allows for efficient deployment compared to larger generative models.
  • Developers using Jev Clients: Direct compatibility with existing Jev infrastructure.

Finetunes
5 models