Qwen-Image-2.1 runs on a Mac with 16 GB, but slowly.

Running Qwen-Image-2.1 on a MacBook Air with 16 GB of memory is possible, but it does not promise a comfortable experience. In the test, generating a 1024 × 1024 image took about 45 seconds per step, and the system used memory compression with swapping to the SSD.

The catch is the peak MLX memory figure of 18.33 GB. The MacBook physically has 16 GB, but macOS can compress data and swap it to the SSD. The model continues to work, though the cost of such a setup is measured in time. Interactively refining prompts will be painful, but queuing tasks and coming back later is quite realistic.

For Apple Silicon, the article recommends an MLX build; for other systems, there is a GGUF option with ComfyUI. On a Mac, the --low-ram mode is especially important: it unloads the text encoder after processing the prompt and uses tiled VAE decoding. In the test on an M1 Max, peak memory at 768 × 768 dropped from 33–36 GB to approximately 7 GB. This is a different computer, but it clearly shows how much the memory-saving mode changes the picture.

There is also a practical caveat for those who care about rights to the output. Qwen-Image-2.1 is distributed under the Qwen Research License, which restricts commercial use. Distilled models that require 4–8 steps are available for faster generation, but the author notes a loss of quality in complex editing and compositions with multiple references.

The test has significant limitations. Generation was tested, but editing with multiple references was not tested on a 16 GB Mac. In the 4-bit result, the texture of the wool was repeated, distant signs were illegible, and the architecture lost precision when enlarged. A build without a safety checker also leaves the user with more responsibility for the content being generated.

Would you wait for generation on a Mac like this for the sake of privacy and local control, or choose a cloud service for the sake of speed?

Source: All Articles / Machine Learning / Habr