Image generation

12 aliases · FLUX.2 klein · Z-Image-Turbo · Qwen-Image · FLUX.1-schnell · HiDream-O1 · Stable Diffusion XL / 3.5 · Bonsai Image. Local image generation through an OpenAI-compatible /v1/images API — the same server the desktop app's Images tab uses.

Pick one

One command per model, smallest memory floor first. The floor is the alias's minimum-memory setting; the size is what rapid-mlx models --modality image-gen reports. Install the extra once with pip install rapid-mlx==0.16.0 (Python 3.11+); weights download on first use.

• rapid-mlx serve qwen-image-2.1
  Qwen-Image 2.1, mflux 4-bit · 8.9 GiB download · 8 GB+ Mac · generate + edit
• rapid-mlx serve flux2-klein-4b
  FLUX.2 klein 4B, 4-bit — the fast default · 4.3 GiB download · 12 GB+ Mac · generate + edit
• rapid-mlx serve bonsai-image-4b-2bit
  ternary FLUX.2 klein 4B, 2-bit · 3.6 GiB download · 12 GB+ Mac · generate
• rapid-mlx serve z-image-turbo
  Z-Image-Turbo, 4-bit, 8 steps · 5.5 GiB download · 16 GB+ Mac · generate
• rapid-mlx serve flux-schnell
  FLUX.1-schnell 12B, 4-bit · 9.0 GiB download · 16 GB+ Mac · generate
• rapid-mlx serve sdxl-base
  Stable Diffusion XL Base 1.0 · 6.5 GiB download · 16 GB+ Mac · generate
• rapid-mlx serve flux2-klein-4b-bf16
  FLUX.2 klein 4B, full precision · 14.9 GiB download · 32 GB+ Mac · generate + edit
• rapid-mlx serve hidream-o1-dev
  HiDream-O1 Image Dev · 16.4 GiB download · 32 GB+ Mac · generate
• rapid-mlx serve sd35-large-4bit
  Stable Diffusion 3.5 Large, 4-bit · 15.3 GiB download · 32 GB+ Mac · generate
• rapid-mlx serve qwen-image-2.1-bf16
  Qwen-Image 2.1, full precision · 30.9 GiB download · 32 GB+ Mac · generate + edit
• rapid-mlx serve qwen-image
  Qwen-Image, mflux q6 · 28.9 GiB download · 64 GB+ Mac · generate
• rapid-mlx serve qwen-image-edit
  Qwen-Image-Edit, mflux q8 · 34.9 GiB download · 96 GB+ Mac · edit only

rapid-mlx serves POST /v1/images/generations (and POST /v1/images/edits) in the OpenAI Images shape, rendering through mflux and pinned native MLX runtimes entirely on your Mac. One image model is resident at a time — same single-worker discipline as the video lane. In the desktop app, the Images tab drives the same endpoints: pick a model, prompt, refine, and every render lands in a filmstrip you can step back through.

family

Image generation (mflux + native MLX)

aliases

12

install

pip install rapid-mlx==0.16.0

OpenAI base URL

http://localhost:8000/v1

Usage

pip install rapid-mlx==0.16.0
rapid-mlx serve z-image-turbo

curl http://localhost:8000/v1/images/generations \
  -H 'Content-Type: application/json' \
  -d '{"model":"z-image-turbo","prompt":"a lighthouse at dusk, oil painting","size":"1024x1024"}'

Family pages

Z-Image, FLUX and Qwen-Image have their own pages with usage and download details:

Notes & caveats

• Requests are validated against the served model — a mismatched model field returns a clear error instead of silently ignoring it.
• Generation is compute-bound: seconds to tens of seconds per image depending on model and Mac. Distilled checkpoints (FLUX.2 klein, Z-Image-Turbo, FLUX.1-schnell) take 4–8 steps; the others take 20–40.
• Licenses differ by model: HiDream-O1 is MIT, SDXL is OpenRAIL++, Stable Diffusion 3.5 Large uses the Stability AI Community License — check a model's terms before commercial use.
• The desktop app keeps a chat model and an image model warm at once, so switching tabs doesn't evict either.

Frequently asked questions

Can I generate images locally on a Mac?

Yes — POST /v1/images/generations renders entirely on Apple Silicon. No cloud, no per-image cost. The desktop Images tab uses the same server.

Which image models run on Apple Silicon?

Twelve aliases, from qwen-image-2.1 and flux2-klein-4b (8–12 GB Macs) up to qwen-image-edit (96 GB). flux2-klein-4b is the fast default.

Where next