Evidence
Controlled comparisons
A-versus-B questions the claim registry can answer honestly: both sides measured on the same unit under the same frozen workload, one variable changed. Every number links to its permanent claim page with run evidence.
- llama.cpp Vulkan vs ROCm on AMD Strix Halo
Which llama.cpp backend should you run on a Ryzen AI Max+ 395 (gfx1151) box for single-user chat — Vulkan or ROCm?
Vulkan · ROCm
- Q4_K_M vs Q8_0: does the bigger quant buy anything?
Is it worth spending the extra memory on the Q8_0 artifact of Qwen3.6-35B-A3B for chat and JSON automation on Strix Halo?
Q4_K_M (20 GB) · Q8_0 (35 GB)
- Thinking vs no-thinking: what reasoning costs in a chat
Should you leave the reasoning block on for everyday assistant tasks on a local Qwen3.6-35B-A3B?
Reasoning off · Reasoning on (4096)
- Prompt cache on vs off: the second question over the same document
Does the second question over the same pasted document pay the full prefill again on a local llama.cpp server?
Cache on · Cache off