Local coding and homelab on a first-gen AMD Framework 13 (Ryzen 7 7840U)

On a first-gen AMD Framework 13 (Ryzen 7 7840U, 64GB of unified memory, 2TB SSD), local work is driven by a stated need to write application and devops code without SaaS models, for privacy and company policy, plus heavy homelab use. Two models run through llama.cpp (build 9190) with the hermes harness at a 128K configured context and two concurrent requests: Qwen 3.6 35B-A3B (Q4_K_XL) at about 24 tok/s decode and 306 tok/s prefill, and Qwen 3.8 Flash Next (UD-IQ3_XXS) at about 6 tok/s decode and 88 tok/s prefill. The larger model is too heavy for the machine, usable only by leaving a task running and returning later, and llama.cpp occasionally stops serving tokens until it is restarted; the 35B MoE runs well, and hermes lets the model notice and correct its own mistakes.