From GitHub · · 1m 40s
Kev: Small Decision Models for Self-Training
Train and run Jev-like decision models on your own with Kev.
In brief
Train and run Jev-like decision models on your own with Kev. Kev is a family of small decision models built on Qwen3.5/3.8 that you can train and run locally for customer support routing. Originally reported by GitHub.
Quick answers
- What is Kev?
- Kev is a family of small decision models built on Qwen3.5 and Qwen3.8.
- What is the key takeaway from Kev: Small Decision Models for Self-Training?
- Kev is a family of small decision models built on Qwen3.5/3.8 that you can train and run locally for customer support routing.
What is Kev?
Kev is a family of small decision models built on Qwen3.5 and Qwen3.8.
Key Features
Yes/no, multiple-choice, and rating questions in one request with calibrated probabilities.
Drop-in for Jev
The TypeSafe Python SDK works against Kev servers unchanged.
Model Sizes
Four sizes from 0.8B for laptops to 27B for data-center GPUs.
Accuracy
Kev-27B matches Jev on new sources with 0.848 vs 0.857 accuracy.
Local Deployment
Run Kev-4B locally with one command using Python and uv.
Fine-Tuning
Fine-tune on your own labeled examples with a coding agent.
Deployment
Deploy your own HTTPS endpoint with zero cost when idle.
Use Case
Route customer support tickets with department, urgency, and frustration scores.
Model Performance
Kev-4B answers six questions in 18.1 ms on an H100 GPU.
Training
Train on public datasets and fine-tune with one epoch on 5,219 examples.
Modal Deployment
Use Modal to deploy Kev with GPU scaling and zero cost when idle.
Accuracy Comparison
Kev-27B trails Jev on knowledge questions but matches on new sources.
Limitations
Calibration and date arithmetic are weaker than Jev for customer support.
Self-Training
Kev learns from your data without needing a large dataset.
What's Next
Kev will improve with more fine-tuning on customer support data.
The short version
Kev is a family of small decision models built on Qwen3.5/3.8 that you can train and run locally for customer support routing.
