Putting Kimi Delta Attention into a Qwen model
Gated DeltaNet is KDA with one knob held still, so the swap costs no distillation and no training tokens — verified at 0.77× the kernel-noise floor. The weight mapping, the rank-16 construction that makes it work, five traps that cost us real time, and what the conversion turns out to be worth once you have it.
Read →A small detector that doesn't cry wolf
We shipped the first Jataayu model — a 0.8B prompt-injection detector, Apache-2.0, on the Hub. The hard part wasn't catch-rate; it was over-defense. The data fix cut benign false positives from 82 to 6 while recall held at 0.85.
Read →The Harness Is the Point
A cited, 45-paper survey of harness engineering at ICML 2026 — how the interesting design surface moved from the model to the runtime, the tool contract, the orchestrator, and the effect boundary.
Read →How Do We Make It Realize Something Went Wrong?
Why closing the self-improvement loop is so much harder than drawing it — the mechanics are easy; the rater, the part that knows a wrong answer was wrong, is the whole game.
Read →