Qwen3.6-35B-A3B hits 21.7 tok/s on consumer GPUs

Sparse MoE architectures are making high-end reasoning viable on consumer-grade setups, though prompt processing remains a significant bottleneck compared to dense models.

• –Hybrid offloading (--cpu-moe) provides "free" performance gains by offloading inactive experts to system RAM without sacrificing generation speed.
• –The model shows a major reasoning leap, outperforming the Qwen 3.5 dense variant by a substantial margin in agentic benchmarks like Terminal-Bench 2.0.
• –PCIe bandwidth limits ingestion efficiency, leaving dense models with a nearly 2x advantage in prompt processing speeds.
• –Technical stability remains a challenge; current custom llama.cpp builds crash when combining Gated Delta Net optimizations with hybrid offloading.
• –Real-world tests confirm autonomous reliability, with the model successfully completing multi-step tool calls for infrastructure automation.