Qwen3.6-35B-A3B hits 130 t/s on consumer GPUs

Qwen3.6-35B-A3B's MoE architecture is a major milestone for local AI, providing elite coding capability at speeds previously reserved for much smaller models. The sparse MoE with only 3B active parameters per token enables lightning-fast inference while maintaining 35B-class reasoning. Currently, IQ4 quantization offers the optimal tradeoff between speed and reasoning accuracy for local hardware, with performance gains of 10-15 t/s possible using specialized coding presets from Unsloth. Its 262K native context window and 73.4% SWE-bench score position it as a formidable local competitor to cloud models.