I picked up the AMD Strix Halo with 128GB a year ago, and had my company pick up a framework one prior to that.

Qwen3.8-Flash-Next is the equalizer. I'm building autonmously with it for personal and work. There's a halogen-flash-server that runs rings around llamacpp.

With the right engineering harness, I'm keeping the same agent busy through 6 million tokens by recycling the context with dynamic context pruning and eviction so it's able to just keep chugging with a little direction & decisions.