jaredpalmer/kev

What it solves

Kev is designed to replace hosted decision models like Jev, allowing users to run small, high-performance decision models locally or on their own infrastructure. It solves the problem of needing a reliable way to perform multiple-choice, yes/no, and rating tasks on a single piece of text without relying on a proprietary API, while providing calibrated probabilities for confidence-based routing.

How it works

Kev is a family of models based on the Qwen architecture (Qwen3.5 and Qwen3.8), implementing a decision-model architecture. It allows users to ask multiple questions (yes/no, multiple-choice, or scores) about a document in a single request. The system provides calibrated probabilities for each answer, meaning the confidence scores are reliable for automating decisions.

Who it’s for

It is intended for developers and organizations that need to automate decision-making tasks such as ticket routing, escalation rules, and sentiment analysis, and who prefer to self-host their models for privacy, cost, or performance reasons.

Highlights

• Multiple Question Types: Supports yes/no (noul), multiple-choice (choice), and rating (score) questions in one request.
• Calibrated Probabilities: Each model checkpoint includes a fitted temperature to ensure confidence scores are reliable.
• es-drop-in Replacement: API compatibility with TypeSafe's System One, allowing existing Jev users to switch seamlessly.
• Flexible Deployment: Available in four sizes (0.8B to 27B parameters) and runs on CUDA, ROCm, or Apple Silicon (MLX).
• Custom Fine-tuning: Users can fine-tune models on their own labeled examples using provided scripts or a coding agent skill for Modal.
• Large Context Window: Supports documents up to 65,536 tokens, with the Kev-27B model maintaining accuracy at this length.