What this article covers
From building ZumiForge. Get ZumiForge, free for Mac.
ZumiForge is our Mac app for making sets of images and videos. It can send a prompt to models running on the Mac itself, or to cloud services using your own keys. Since May it has run 2,444 jobs for us: course thumbnails, storyboards, product shots and a lot of experiments.
That job history is a useful record. Every job keeps its model, size, steps and status, and recent ones keep how long they took. So rather than guess whether local models are "good enough", this article answers from what actually happened.
The machine and the models
All the timings below come from one Mac: an Apple M5 Max with 128 GB of memory. That is the top of the range, and it matters. Local image models live in memory, and a large model on a smaller Mac will run slower or not load at all. We haven't measured smaller machines yet, so treat these numbers as a ceiling, not a promise.
The local models run through mflux, an implementation of image models on Apple's MLX framework, plus separate runtimes for MiniMax H3 Image and LTX-2.5 video. Nothing in a local job leaves the computer.
How fast is local?
These are medians from the jobs where ZumiForge recorded a duration:
| Model (on the Mac) | Size | Steps | Timed jobs | Median | Slowest 10% |
|---|---|---|---|---|---|
| MiniMax H3 Image | 1376 × 768 | 6 | 29 | 13.6 s | 16.5 s |
| Z-Image Turbo | 1360 × 768 | 9 | 104 | 22.2 s | 25.7 s |
| Z-Image Turbo | 1920 × 1080 | 9 | 9 | 37.8 s | 40.9 s |
| Qwen-Image 2.1 Turbo | 1920 × 1080 | 8 | 10 | 46.9 s | 50.3 s |
| LTX-2.5 (video) | 640 × 512 | 20 | 13 | 53.8 s | 167 s |
At around 20 seconds an image, a set of eight 16:9 thumbnails takes about three minutes. Small preview sizes (282 × 226) come back in 4 to 10 seconds on most models, which makes them good for trying a prompt before running the full set.
For comparison, the cloud jobs we timed through Runware for Ideogram at 1280 × 720 came back in about 12 seconds. The cloud is still faster per image. But a local job costs nothing, and it isn't competing with anyone else's queue.
Where local is good enough
The clearest sign is what we actually use. Z-Image Turbo, running locally, is our second most-used model at 892 jobs, and it made every thumbnail for our newest course series. In practice local models do well at:
- Illustration in a fixed style. Describe the style once ("layered paper-cut illustration, deep teal, coral and mustard on cream") and every image in a set follows it.
- Iteration. When an image costs nothing, you try five versions of a prompt instead of settling for the first.
- Large batches. Queue a set and leave it. There's no per-image bill growing in the background.
- Anything private. Unreleased product shots, client work or personal photos used as references never go to a third party.
Where it still falls short
Our most-used model overall isn't local. It's Google's Nano Banana Pro, at 991 jobs, through the cloud with our own key. The reasons are the honest list of what local models can't yet do as well:
- Real text in the image. If a picture needs a readable word on a sign or a screen, the best cloud models are far more reliable. Local models tend to produce convincing gibberish.
- Following complex instructions. A prompt with several people, specific positions and a particular action is more often right first time in the large cloud models.
- Editing with several references. Combining a character from one image with a place from another works locally with H3 Image, but the cloud editing models are more consistent.
- Video. LTX-2.5 runs locally, but clips are short and small, and the slowest 10% took nearly three minutes each. For anything longer than a few seconds, the cloud video models are in a different class.
Failures are real in both places. 77 of the 2,444 jobs ended in an error, about 3%, spread across local and cloud. Usually it was a model that wouldn't load, a service limit or a rejected prompt. ZumiForge keeps the exact job so it can be retried as it was.
Choosing per job, not per tool
The useful question isn't "local or cloud?" but "which jobs need the cloud?" Our working rule:
- Try the prompt locally at preview size.
- If the style and composition are right, run the full set locally.
- Send only the images that need what local can't do, such as legible text, a hard composition or long video, to a cloud model.
Because ZumiForge can send the same prompt to several models at once, step 1 often includes one cloud model for comparison. Seeing four versions side by side settles the question faster than any benchmark.
What I would tell someone starting out
Memory is the spec that matters. More memory means larger models load and run without swapping.
Use preview sizes to iterate. A 5-second preview tells you if the prompt works before you spend 20 seconds an image on the full set.
Keep a job record. Model, size, steps, prompt and time per job is what made this article possible, and what lets you rerun a good result exactly.
Don't ask local models for words. Design the picture without text, and add any words in the page or layout instead.
Queue a whole set of images, on your Mac
Free models run on Apple silicon and nothing leaves your computer. Cloud models run with your own keys, kept in the Keychain. No account.
Get ZumiForge