Qwen2.5-Coder 3B

User reports Qwen2.5-Coder 3B running at roughly 4 to 5 tokens per second on a Snapdragon 870 Android tablet with 8 GB of RAM, using Ollama under Termux. The pipeline loads one model at a time and unloads it with ollama stop, waiting 10 seconds before the next, which allows chains of 6 or 7 models without Android killing Termux. Three critic agents were removed to keep the pipeline stable on the tablet. The user asks whether others have tried similar sequential loading or memory management on low-end hardware.