Jared Palmer
Short summary
Jared Palmer released Kev, a family of four open-source models (0.8B to 27B parameters) designed to answer multiple-choice and yes/no questions by scoring answer options directly rather than generating tokens. The models run on consumer hardware, come with Apache-2.0 weights, and match or exceed the performance of TypeSafe's proprietary Jev API.
- Kev-27B achieves 75.7% accuracy on generalization tasks and 91.8% on policy reasoning, with smaller variants (Kev-4B, Kev-9B) performing within 1-2 percentage points on specific workflows while using far less compute.
- The architecture uses a pointer head to score supplied answers without token generation, and caches document computation across multiple questions—cutting inference time from 9.4 seconds for a new 64k-token document to 0.7 seconds per follow-up question on an H200.
- Kev-4B runs on a 32GB Mac (136ms per question with cached text) and Kev-0.8B on any Apple Silicon, making local deployment and fine-tuning practical; the largest model costs $44.09 per million requests on H200s versus $3.54 for the smallest.
Full summary
Jared Palmer released Kev 1.0, a family of four open-source decision models ranging from 0.8B to 27B parameters that work like TypeSafe's Jev but with weights you can run and fine-tune yourself. The models take text, questions, and answer options, then return probabilities for each option. All four models are available on Hugging Face under Apache-2.0 license, with code on GitHub that includes a server running on GPUs and Apple Silicon via MLX. You can test Kev-4B in a browser or deploy your own version.
The performance varies by model size and task. Kev-27B leads overall accuracy, hitting 91.8% on policy reasoning and 90.8% on sorting consumer complaints, though it requires expensive hardware like an H200 GPU. Kev-9B and Kev-4B come within a percentage point on consumer complaints and match Kev-27B on developer-tool decisions, making them practical alternatives. Kev-4B runs on a 32 GB Mac, answering five questions about cached text in 136 milliseconds. The evaluation covers six panels including generalization tests on 14 public datasets and held-out examples from trained task families, though Palmer notes these are his own measurements, not independent benchmarks.
The architecture uses a pointer head that scores answer options directly rather than generating tokens, which saves computation when you ask multiple questions about the same document—the text gets read once and cached. Kev-27B trained on about 146,000 examples and 337,000 questions, combining existing data with licensed public tasks and synthetic examples. Palmer trained the full backbone on eight H200s for roughly 16 hours at about $650 in GPU costs. The new Kev-9B received an additional three-hour fine-tuning pass that closed performance gaps, jumping from 58% to 83% on policy reasoning and 64% to 79% on developer tools.