do small edge boxes make sense with llama.cpp

[tensorport]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there.