Qwen 3.6 just dropped, and it is a masterclass in sparse Mixture-of-Experts (MoE) efficiency. If you have been looking for a local model that actually handles repository-level coding without melting your VRAM or your patience, this is the one.<\/p>\n\n\n\n
On April 2, 2026, Alibaba Cloud officially released the Qwen 3.6 series<\/a>, marking a significant pivot toward “agentic” AI\u2014models designed not just to chat, but to use tools, execute code, and manage multi-step plans. The standout for practitioners is the Qwen3.6-35B-A3B<\/a>, a sparse MoE model that boasts 35 billion total parameters but only activates 3 billion parameters per token<\/strong> during inference.<\/p>\n\n\n\n The performance numbers are aggressive. It scores 73.4% on SWE-bench Verified<\/strong>, putting it in direct competition with frontier models like Claude 3.5 Sonnet for autonomous software engineering tasks. In reasoning benchmarks, it hits a 92.7 on AIME 2026<\/strong> and 86.0 on GPQA<\/strong>, suggesting that the “thinking mode” introduced in this version is doing heavy lifting. While the flagship Qwen3.6-Plus remains behind an API, the 35B-A3B weights have been released under the Apache 2.0 license<\/a>, making it a prime candidate for local deployment.<\/p>\n\n\n\n The architecture of Qwen 3.6-35B-A3B is a sophisticated hybrid designed to solve the “long context vs. compute cost” trade-off. It utilizes a causal language model structure with a unique vision encoder layout described as Key technical specifications include:<\/p>\n\n\n\n To run Qwen 3.6-35B-A3B locally, The Quick Test<\/strong>: Run the command above with a complex coding prompt. You should see the model generate a Qwen 3.6 enters a crowded field but carves out a specific niche for “Agentic Workflows.”<\/p>\n\n\n\n For developers building IDE extensions or autonomous agents, Qwen 3.6 is the current “best-in-class” for performance-per-watt and performance-per-VRAM-GB.<\/p>\n\n\n\nUnder the hood<\/h2>\n\n\n\n
10 \u00d7 (3 \u00d7 (Gated DeltaNet \u2192 MoE) \u2192 1 \u00d7 (Gated Attention \u2192 MoE))<\/code>.<\/p>\n\n\n\n\n
preserve_thinking<\/code> parameter allows the model to retain reasoning context (Chain-of-Thought) across conversation turns. This prevents the common issue where a model “forgets” its logic during a multi-turn debugging session.<\/li>\nHow to try it yourself<\/h2>\n\n\n\n
llama.cpp<\/code> is currently the most optimized path, especially if you want to leverage the MoE offloading features.<\/p>\n\n\n\nPrerequisites<\/h3>\n\n\n\n
\n
cmake<\/code>, gcc<\/code> or clang<\/code>, and the latest llama.cpp<\/code> build.<\/li>\n<\/ul>\n\n\n\nSetup and Install<\/h3>\n\n\n\n
\n# Clone and build llama.cpp\ngit clone https:\/\/github.com\/ggml-org\/llama.cpp && cd llama.cpp\ncmake -B build && cmake --build build --config Release\n\n# Download the GGUF (example using huggingface-cli)\nhuggingface-cli download Qwen\/Qwen3.6-35B-A3B-GGUF qwen3.6-35b-a3b-q4_k_m.gguf --local-dir .\/models\n<\/code><\/pre>\n\n\n\nRunning with MoE Optimization<\/h3>\n\n\n\n
--n-cpu-moe<\/code> flag is the secret sauce here. It allows you to offload expert components to the CPU if you are VRAM-constrained, though for maximum speed on a card like the RTX 5070 Ti, you should set it to 0 to keep everything on the GPU.<\/p>\n\n\n\n\n.\/build\/bin\/llama-cli -m .\/models\/qwen3.6-35b-a3b-q4_k_m.gguf \\\n -n 512 --ctx-size 8192 \\\n --n-gpu-layers 81 \\\n --n-cpu-moe 0 \\\n --chat-template-kwargs '{\"preserve_thinking\": true}' \\\n -p \"You are a senior staff engineer. Review this Python function for race conditions: ...\"\n<\/code><\/pre>\n\n\n\n<thinking><\/code> block before the actual code. If you are getting gibberish, ensure your context length is at least 8192 and check that you aren’t using CUDA 13.2, which has known compatibility issues<\/a> with this release.<\/p>\n\n\n\nWhere this fits<\/h2>\n\n\n\n
\n
What practitioners are saying<\/h2>\n\n\n\n