Questions
Short, self-contained answers to the questions people actually search. Each page opens with a Problem / Claim / Conclusion lede a non-specialist can absorb in thirty seconds, then synthesises the record behind it — every number is a claim chip, and the evidence list at the foot shows each cited claim's current state.
Does KV quantisation cost quality?
Mostly no — once a one-commit engine patch is applied. The patch recovers +70.3% decode speed at production depth clm-0022, makes quantised KV the cheapest arm in electricity on fixed-length work — 42% less energy clm-0041 — and shrinks the agentic cost from +39% extra conversation turns clm-0038 to +9.3% clm-0045. The memory saving comes from quantisation itself; the patch removes most of what it used to cost.
Vulkan or ROCm on Strix Halo?
On this chip: Vulkan, in every cell measured so far. At f16 KV, stock Vulkan beats stock ROCm at every depth from 0 to 131,072 tokens — decode +19-21% at every depth, prefill +4% to +20% growing with depth clm-0050. With quantised KV a community fork becomes the best long-context arm measured: 42.5 tok/s decode at 131,072, 2.3x stock ROCm clm-0051. One model measured so far — and a pending BF16 patch shows the lead can invert with the right kernel work clm-0046 — so a backend verdict is a depth-and-build statement, never a permanent truth.
Is thinking mode worth it?
Per-model, not global. On the lab's production 122B, thinking is a net negative: OFF solves 5/5 airline tasks where ON solves 3/5 and deadlocks indefinitely on the rest, with identical reward on every task both arms complete clm-0033. A cross-model matrix suggested thinking helps other models — but those score gaps fell inside the harness's own noise and were retracted clm-0035 clm-0036. What survives is categorical: the lever flips behaviour per model, so it has to be measured per model.