Qwen 3.6 is the latest model in the Qwen series, released in June 2026. It features a dual lineup strategy, offering both a 27B dense and a 35B-A3B MoE (active 3B) architecture simultaneously. By releasing models of the same generation in both dense and MoE formats, users can choose the optimal version based on the task characteristics and hardware environment. In terms of performance, the 27B dense model achieves a level close to the initial version of Claude 3.5 Sonnet and ranks among the top open-source 27B models as of June 2026. The 35B-A3B MoE version has a total of 35B parameters, but only 3B are actually activated, allowing it to run with extremely low VRAM (approximately 3-4GB). This makes it runnable on laptop GPUs or integrated graphics, drastically improving accessibility, almost like having a high-performance AI that you can carry around in your pocket. Of course, since the active parameter is 3B, the quality is lower compared to the 27B dense model, but it still demonstrates sufficient performance for simple classification, summarization, and translation tasks. The 27B dense model can run with approximately 17GB of VRAM when using Q4 quantization, and even with BF16 full precision, it can be run on two consumer GPUs. Ollama supports various quantization options, including Q4/Q8/BF16, allowing for fine-grained adjustments to the trade-off between precision and speed. It is licensed under Apache 2.0, allowing for commercial use, and it also inherits the thinking mode of Qwen 3/3.5, enabling step-by-step reasoning in complex inference tasks.