Qwen3.5-35B-A3B tops 27B dense for long context

The hybrid Gated DeltaNet and MoE architecture enables linear compute scaling, making 262K context performant on a single 3090 or 4090. By activating only 3B parameters, the model achieves a 3-4x speedup over dense models like Qwen 2.5 32B. IQ4XS quantization is the recommended sweet spot to avoid performance collapse, while the native 262K window (extensible to 1M) avoids common needle-in-a-haystack issues.