Local image generation on an M4 Max: FLUX.2 Klein 4B and Qwen-Image
- Date
- 8 March 2026
- Task type
- Image generation
- Status
- Complete
Models tested
| Model or tool | Provider | Open? | How run |
|---|---|---|---|
| FLUX.2 Klein 4B (4B parameters, 8-bit, run with mflux) | Black Forest Labs | Open weights | Local |
| Qwen-Image 2512 (20B parameters, 4-bit MLX build, Apache 2.0, run with mflux 0.16 or later) | Alibaba | Open weights | Local |
Purpose
Find out how long Qwen-Image takes to generate an image on Apple Silicon, and how that compares with FLUX Klein.
Context
An Apple M4 Max computer with 128 GB of unified memory and 40 GPU cores. Both models were run through the mflux command-line tool, one image per setting, in a hands-on session.
Method
Images were generated at the resolutions and step counts in the table and total time was noted for each.
Results
| Model | Quantisation | Resolution | Steps | Time |
|---|---|---|---|---|
| FLUX.2 Klein 4B | 8-bit | 512 x 512 | 4 | about 7 s |
| FLUX.2 Klein 4B | 8-bit | 1024 x 1024 | 4 | about 18 s |
| Qwen-Image 2512 | 4-bit | 1024 x 1024 | 20 | about 8 min |
| Qwen-Image 2512 | 4-bit | 1328 x 1328 | 20 | about 10 min |
Caveats
- One machine, one run per setting (n=1), timings approximate.
- Image quality was not scored; only generation time is compared here.
- The two models were run at different quantisation levels (8-bit and 4-bit).
- An earlier estimate of about 3 minutes for Qwen-Image was too low.
What this suggests
On this machine Qwen-Image ran at about 23.5 s per step, which made a 20-step image take about 8 minutes against about 18 s for FLUX Klein at 1024 x 1024. The notes record that Qwen-Image renders text in images better, but that was not measured here.
Source
entries/2026-03-08-local-image-generation-m4-max.md in the GitHub repository (Markdown with YAML frontmatter).