Qwen-Image-2.1 Benchmark on Apple Silicon: ~1.78s per Step via MLX bf16

Qwen Developers 1 hour ago

Key Info

Qwen-Image-2.1 was benchmarked locally on an Apple M5 Max using MLX in bf16 (via a custom MFLUX pipeline), generating a 1600×672 image in about 26s at 10 steps and 97s at 50 steps. The run averaged roughly 1.78s per diffusion step plus ~8s of fixed overhead, with peak memory around 31GB.

Highlights

At 1600×672 (2.39:1) with a fixed seed, 10/20/30/40/50 steps took about 26/44/61/79/97 seconds respectively.
Per-step cost was ~1.78s, with ~8s one-time overhead for loading, prompt encoding, and VAE decode.
Peak memory was ~31GB in bf16, giving a practical data point for running Qwen-Image-2.1 on high-end Apple Silicon setups.

Original post

worth looking at

Quoted post

Qwen-Image-2.1 experiments on Apple Silicon (M5 Max, MLX, bf16) with my custom MFLUX 1600×672 (2.39:1), fixed seed: 10 steps → ~26s 20 steps → ~44s 30 steps → ~61s 40 steps → ~79s 50 steps → ~97s ~1.78s per step + ~8s fixed (load, prompt encode, VAE decode) Peak memory: 31 GB Prompt: A city street bends upward at the horizon: the ground and its buildings curve up from street level and arc into the sky like a giant wave, one single continuous landscape where skyscrapers grow progressively more vertical until they tower overhead, the road itself rising and wrapping up into the clouds. Dream physics like a folded dream city: no mirror duplication, just one seamless city rolling from horizontal to vertical. Wet reflective asphalt, cool blue-grey palette with warm window lights, glass and steel towers, low clouds threading between the rising districts, slight lens distortion, IMAX 70mm film look, cinematic wide-angle, photorealistic, volumetric haze, epic impossible scale.

#Qwen-Image-2.1#Apple Silicon#MLX

View original on X