kirp/jpt-4b

JPT-4B: A Fast, Open Decision Model

JPT-4B is a 4.5 billion parameter model developed by kirp, fine-tuned from Qwen/Qwen3.5-4B. It is engineered as an open decision model that provides calibrated probabilities for various question types from a single forward pass, prioritizing speed and direct answers over generated explanations. This model implements the typed-decision interface, accepting choice, score, and noul (yes/no) questions.

Key Capabilities & Differentiators

• Rapid Decision-Making: Designed for low-latency applications, it provides probabilistic answers without generating reasoning tokens, making it highly efficient.
• High Accuracy on Decision Benchmarks: JPT-4B achieves the highest public accuracy (0.879) on JevBench v1.4.0 among all systems, and scores 43.04 on Decision Index 0.2.1, outperforming other 4B entrants.
• Multimodal Input: Built on Qwen3.5-4B, its vision tower remains unchanged, allowing it to process both text and image inputs (screenshots, photos, video frames) for decision-making.
• Fine-tuned Performance: Through LoRA fine-tuning with a Brier loss on 49,221 typed questions, JPT-4B significantly outperforms its base model, Qwen3.5-4B, across various text benchmarks while maintaining image processing capabilities.

Use Cases & Limitations

JPT-4B is ideal for applications requiring fast, probabilistic answers to structured questions, such as automated agents, game environments, or systems needing quick, calibrated confidence scores. It supports up to 255 options for choice questions, though training covered up to 77. While it handles images zero-shot, its weakest areas include arithmetic, dates, and complex belief reasoning (e.g., Minesweeper-style puzzles), as it answers based on evidence without an explicit reasoning phase. It is primarily English-first, with other languages covered by limited multilingual datasets.