Kev: open 'decision models' on Qwen3.5 you can train for $95
Typed questions in, calibrated probabilities out in one forward pass — no prose. Open weights and recipes, runs on a Mac, and it still trails hosted Jev.

Copy markdown
What a 'decision model' actually does
Instead of coaxing an LLM into clean JSON, Kev reads your state and answers many typed questions — yes/no, multiple-choice, 1-5 scores — in a single forward pass, returning calibrated probabilities. The repo's demo routes a support ticket to returns/shipping/billing, flags whether to escalate, and scores frustration, all in one call.
Open all the way down, for ~$95
Apache-2.0, with weights, training recipes and eval data on GitHub and Hugging Face. Palmer says the whole Qwen3.5 port — probes, trials, benchmarks, ablations — cost about $95 in Modal H100 time plus $0.03 of Jev API calls. Each model is just a rank-16 LoRA adapter and a pointer head on a frozen Qwen base.
Trains and runs on hardware you own
Three sizes ship: 0.8B, 4B and 9B. The 4B and 9B serve locally on a 32GB Apple Silicon Mac in bf16; the small one trains in under two hours on an M5, the larger ones in 40-70 minutes on a rented H100. You need transformers 5.17 or newer.
The catch: a preview that still trails Jev
Kev-9B lands around 0.85 out-of-domain against hosted Jev's 0.857, and slips further on multi-step policy reasoning, so Palmer keeps the family unversioned as a preview. Apple Silicon is slow without DeltaNet MPS kernels (~0.8s per five-question request). Strong for routing and scoring today, not yet a drop-in Jev replacement.