Kev: Tiny Jev-like family of decision models built on top of Qwen3.5
Kev introduces an open-source family of Jev-like decision models built on Qwen3.5, providing compact, trainable, and locally runnable AI for classification tasks. This project is gaining traction on Hacker News by offering a transparent alternative to proprietary solutions, enabling developers to integrate efficient decision-making AI directly into their applications. Its emphasis on local deployment and API compatibility with existing tools makes it particularly appealing for practical, real-world use cases.
The Lowdown
Kev is an open-source project offering a family of small, Jev-like decision models built on Qwen3.5 base models. Designed for efficiency and local deployment, Kev provides a transparent and customizable alternative to proprietary decision AI systems, allowing users to train and run these models themselves.
- Model Variants: Available in 0.8B, 4B, and 9B parameters, each with training code and evaluation data.
- Decision Types: Supports various question formats including yes/no (
noul), multiple-choice (choice), and rating (score) questions within a single request. - Architecture: Utilizes a unique attention mask mechanism to enable questions to process shared input text independently, ensuring question isolation.
- Performance & Compatibility: Runs on CUDA and Apple Silicon, with specific performance considerations for different hardware. Its API is designed to match TypeSafe's System One API.
- Training & Fine-tuning: Provides code for training from scratch and fine-tuning on custom datasets, allowing users to adapt models to specific domains.
- Interactive Tools: Includes a web playground for testing inputs and a chess demo showcasing real-time decision-making.
- Limitations: Acknowledged limitations include uncalibrated probabilities on new sources, potential degradation of base model skills post-training, slower performance on Apple Silicon for Qwen3.5 models, and sensitivity to option order.
By open-sourcing these powerful yet compact decision models, Kev empowers developers with a flexible and locally deployable framework for integrating AI-driven decision processes into their applications, from customer support routing to game AI.
The Gossip
Decoding Decision Models & Diverse Decisions
Commenters are eager to understand how Jev-like models differ from traditional neural network classification, noting the flexibility of output classes that don't require pre-training. The discussion quickly pivots to practical applications, highlighting use cases like automating UI/UX styling, building AI for 2D roguelike games, classifying bash tool calls, and acting as verifiers for larger coding agents. This underscores the appeal of fast, local decision-making for specific, structured tasks.
Performance & Practicality
The project's emphasis on low latency and local execution is a significant draw. Users share experiences with similar models, noting the efficiency of smaller decision models for specific tasks compared to the computational cost of larger general-purpose LLMs. Anecdotes about impressive local speeds on various hardware, such as 220ms per move on a 12B model, reinforce the practical appeal for scenarios demanding quick, on-device decisions.
Open Source vs. Knowledge Cutoffs & Deployment
While celebrating the open-source nature, commenters voice concerns about practical limitations. The absence of tool-calling and the inherent knowledge cutoff in static models are discussed as challenges, suggesting the need for continuous retraining or updates. A recurring question asks why such projects aren't typically packaged as Docker images, indicating a strong desire for simpler and more standardized deployment mechanisms.