jaredpalmer/kev
Jev-like family of decision models built on top of Qwen3.5/3.8 you can train and run on your own
What it solves
Kev provides a family of small, specialized decision models designed to handle structured tasks like classification, scoring, and yes/no decision-making. It acts as an open-source alternative to proprietary decision-model APIs, allowing developers to run high-accuracy, calibrated decision-making logic locally or on their own infrastructure.
How it works
Kev models are built on Qwen base models and optimized for specific decision-making tasks. They accept a text input (the "state") and a set of questions, returning structured answers with associated probabilities. The models are designed to be drop-in compatible with existing TypeSafe SDKs. They support fine-tuning on custom datasets to adapt to specific domain requirements, and include calibrated probabilities to help developers set reliable confidence thresholds for automation.
Who it’s for
This project is for developers and engineers who need to automate decision-making workflows—such as ticket routing, sentiment analysis, or policy enforcement—and want to move away from black-box proprietary APIs toward models they can host, fine-tune, and audit themselves.
Highlights
- Offers four model sizes (0.8B to 27B) to balance accuracy and hardware requirements, from laptops to data-center GPUs.
- Provides calibrated probabilities by default, allowing for reliable automated decision thresholds.
- Fully compatible with the TypeSafe System One API, enabling easy migration from existing proprietary setups.
- Supports long-context documents up to 65,536 tokens.
- Includes a fine-tuning workflow that can be automated via a coding agent on Modal, or performed manually with standard scripts.
- Deployable as an HTTPS endpoint that scales to zero when idle to minimize costs.
Related
- Dispatch
- Project
- Project
- Project
- Project