September 22, 2026AgentsOpen SourceBenchmark

Kev is Jev you can train yourself, and the gap is 0.5 points

Jared Palmer put out Kev, a family of small Jev-like decision models you train and run yourself, and it hit Hacker News at 374 points. Apache 2.0 across the board, including the Qwen3 and Qwen3.5 bases it is built on. github.com/jaredpalmer/kev.

The numbers are the story. On new-source test accuracy: Kev-0.8B gets 0.684, Kev-4B gets 0.837, Kev-9B gets 0.852. Hosted Jev sits at 0.857 on dev. Nine billion parameters running on your own hardware lands within half a point of the service, and the 4B is fifteen points above the 0.8B, which is where the real jump lives. Brier scores tell a harsher version of the same story, 0.237 for Kev-9B against 0.211 for Jev, so Jev is still better calibrated even where accuracy has nearly converged.

The architecture is almost aggressively unglamorous. A rank-16 LoRA adapter and a small pointer head on a Qwen base. That is it. The one clever bit is the packing: questions go into a single token sequence with attention masking that stops them reading each other, so you batch a pile of independent decisions through one forward pass and they stay independent. Models updated September 21 after a second training pass on generated examples.

Zoom out and this is the fourth shipped thing in about a week betting that bounded decisions should not go through a frontier model. Cua's CUA-S1 scores existing UI elements instead of generating coordinates. Jev returns calibrated probabilities instead of prose. Vercel's json-render returns a component id instead of markup. Kev says those probabilities should come from a nine-billion-parameter model on your own box. Nobody coordinated this and they are all removing free-form generation from the critical path.

The uncomfortable question Kev forces is the pricing one. If half a point of accuracy is the entire value of the hosted service, what exactly is anyone paying for, and how long does that survive the next base model release? The answer is probably calibration and the data flywheel rather than raw accuracy, and Jev's better Brier score is the only evidence for that on the table right now. https://github.com/jaredpalmer/kev
← Previous
ai-memory lets you quit Claude Code mid-task and finish it in Codex
Next β†’
CodeMidas turns 3,185 random repos into an RL gym for coding agents
← Back to all articles

Comments

Loading...
>_