Jared Palmer released Kev 1.0, a family of four open source decision models. The models range from 0.8B to 27B parameters. They use the same API as TypeSafe’s Jev for text, questions and answer scoring. All weights are Apache-2.0 licensed on Hugging Face. Code and a serving server are on GitHub. Kev-27B leads in overall evaluations. Kev-9B and Kev-4B match it closely on consumer complaints and developer-tool decisions. Models run on GPUs or Apple Silicon via MLX. Kev-27B costs $44.09 per million GPU requests on an H200. Kev-0.8B costs $3.54 per million GPU requests on an L4. Kev caches document processing for multiple questions. Kev-27B was fine-tuned on 146,000 examples and 337,000 questions. Kev-9B received an update to close accuracy gaps.
What changed
No prior open source family of decision models with this API and fine-tuning workflow existed from this author.
What it unlocks
Running and fine-tuning decision models locally with a shared document cache for multiple questions.
- 4 open source decision models: 0.8B to 27B parameters
- 146,000 training examples, 337,000 questions
- Kev-27B: $44.09 per 1M GPU requests, 29 req/s on H200
- Kev-0.8B: $3.54 per 1M GPU requests, 63 req/s on L4
What you need to act on it
- a GPU or Apple Silicon Mac
- MLX for Apple Silicon support
Sources