Kev open-source decision models scale to 0.6B/4B/8B, trainable in 40 min on one H100
TheMoonMidas · x · 2026-09-21
Jared Palmer released Kev, a family of open-source, Jev-like small decision models in 0.6B, 4B, and 8B sizes, built on Qwen3 with a LoRA adapter plus a small pointer head.
How it works: one forward pass reads a document and answers many typed questions in parallel — document and questions are packed into one sequence with a block-causal mask so questions see the document but not each other; a pointer head softmaxes calibrated probabilities with no autoregressive decoding.
Key numbers and features:
- Out of domain, Kev-8B scores 79.6% vs Jev's 85.7%
- Drop-in TypeSafe System One API — works by changing one baseurl
- Kev-4B serves on a 32 GB Mac in bf16 (300 ms for five questions, 40 ms on an H100); repeated documents hit a KV cache for 2-2.5x speedup
- Apache 2.0; Kev-4B trains in 40 minutes on one H100, Kev-8B in 83 minutes
Code, weights, and evals are on GitHub.
More from coding & agent
- Dev ditches Cursor, says he switched back to VS Code and loves it — Arindam_1729 · 2026-09-21
- Agent system layers are mostly transitional, engineer argues — rakyll · 2026-09-21
- Don't feed full HTML: how an agent fixes internal links without hallucination — cneuralnetwork · 2026-09-21
- Engineer quips: AGI is inevitable once we figure out the YAML stuff — rakyll · 2026-09-21
- One RTX 3090 ran Qwen 27B autonomously for 3 weeks — it shipped working CUDA kernels — skeole · 2026-09-21
- Dev buys a Meta coding subscription for its 'excellent model, crazy quota, low price' — intellectronica · 2026-09-21