Homelabbers debate Qwen3.5, GPT-OSS local inference configurations

The convergence of 48GB VRAM prosumer cards and high-efficiency Mixture-of-Experts (MoE) models has shifted the homelab meta from "just fitting the model" to maximizing tokens per second for agentic workflows. Splitting layers across mismatched cards, such as an RTX Pro 5000 and 5060 Ti, is becoming a standard strategy for running 120B+ parameter MoE models like Qwen3.5-122B at usable speeds. Dense models like Qwen3.5 27B still offer a "sweet spot" for reasoning-heavy tasks where context window and throughput are critical for coding assistants. GPT-OSS-20B has emerged as a preferred tool-use model, reflecting OpenAI's successful entry into the open-weights ecosystem, while the choice between vLLM and llama.cpp highlights ongoing fragmentation in local inference engines.