AMD 395+ w/128B. Qwen3-coder-next running from llama.cpp; opencode seems to have the best depth. It can get to around 80k context before pooping out.