Qwen3.6 runs on 32–64GB Macs

Running an artificial intelligence model locally means the model’s weights sit in your computer’s memory instead of on a cloud server. Qwen’s April releases pushed that setup onto Macs with 32GB to 64GB of unified memory. (qwen.ai 1) (qwen.ai 2) Alibaba’s Qwen team open-sourced Qwen3.6-35B-A3B on April 14, 2026, then followed with Qwen3.6-27B on April 21, 2026. Both are positioned as coding-focused multimodal models that developers can download as open weights. (qwen.ai 1) (qwen.ai 2) (github.com) The 35B-A3B model is a mixture-of-experts system, which is a design that activates only part of the model for each token, like switching on a few specialists instead of the whole office. Qwen says it has 35 billion total parameters but only 3 billion active parameters at a time. (qwen.ai) (huggingface.co) That smaller active footprint is the key to the Mac story. A model that does not need all of its parameters live for every step is easier to fit into Apple’s shared pool of CPU and GPU memory. (qwen.ai) (apple.com) Apple’s current Mac Studio lineup starts at 36GB unified memory on M4 Max, can be configured to 64GB or 128GB on higher M4 Max options, and starts at 96GB on M3 Ultra with options up to 256GB. Older Mac Studio models with M2 Max also shipped in 32GB and 64GB configurations, which helps explain why 32GB-to-64GB Macs are the floor many local-model users talk about. (apple.com) (support.apple.com) Qwen says Qwen3.6-35B-A3B outperforms its predecessor Qwen3.5-35B-A3B and beats dense Qwen3.5-27B on several coding benchmarks, despite the much smaller active parameter count. Qwen’s April 21 post also describes Qwen3.6-27B as a dense 27-billion-parameter multimodal model aimed at the size “the community has been asking for most.” (qwen.ai 1) (qwen.ai 2) The hardware angle is getting tighter at the same time. MacRumors reported on April 6 that many U.S. Mac mini and Mac Studio configurations with higher RAM had slipped to delivery estimates of up to four to five months. (macrumors.com) 9to5Mac reported five days later that some of those higher-memory systems had moved from long waits to “currently unavailable,” including an M4 Mac mini with 32GB RAM, an M4 Pro Mac mini with 64GB RAM, an M4 Max Mac Studio with 128GB RAM, and an M3 Ultra Mac Studio with 256GB RAM. (9to5mac.com) The result is a narrow window for local artificial intelligence buyers. Qwen’s new open-weight models lowered the software barrier just as Apple’s higher-memory desktops became harder to get. (qwen.ai) (qwen.ai) (macrumors.com) That is why “Qwen3.6 runs on 32–64GB Macs” is less about one benchmark than about a stack coming into focus: open weights from Qwen, unified memory from Apple, and a shrinking supply of the machines people want to run them on. (github.com) (apple.com) (9to5mac.com)