Running Qwen3.8-35B-A3B on consumer hardware

A 35B coding model, Qwen3.8-35B-A3B with 3B active, can now run at 50 tokens per second on a $250 12GB RTX 3060 GPU using 16GB RAM and hybrid CPU/GPU offload. The setup achieves 550 tokens per second for prompt processing with a 262K configured context. This performance demonstrates how mixture-of-experts and hybrid offloading can make cloud inference less necessary for certain coding workflows. Open article →