mlximages

Apple Silicon (MLX) builds of Bonsai Image 4B, mirrored for on-device text-to-image generation in Privacy AI. Nothing here is original work: these are redistributed Apache-2.0 models, repackaged so that each one is self-contained.

Directory Weights Download
bonsai-image-4B-ternary-mlx 2-bit (ternary) ~3.97 GB
bonsai-image-4B-binary-mlx 1-bit (binary) ~3.51 GB

Both arms render at 512x512 and 1024x1024 in four steps. They share a byte-identical 4-bit Qwen3 text encoder, which is why the binary arm is smaller to download but needs the same memory to run — the peak is the text encoder, not the transformer.

Layout

bonsai-image-4B-{binary,ternary}-mlx/
  transformer-packed-mflux/     packed diffusion transformer
  text_encoder-mlx-4bit/        Qwen3 4-bit text encoder
  tokenizer/                    includes chat_template.jinja
  vae-small-decoder/            relocated from black-forest-labs/FLUX.2-small-decoder
  LICENSE  NOTICE.md  README.md  MIRROR-NOTES.md

Each arm's MIRROR-NOTES.md records exactly what was removed, what was added and the source repository and revision of each — Apache-2.0 §4(b), since the payload composition differs from upstream even though no file was edited.

Two deliberate differences from the upstream arms:

  • vae/ is removed. The full VAE has different BatchNorm statistics from the small decoder this pipeline actually runs, so shipping it would only invite the wrong one being loaded.
  • vae-small-decoder/ is added, taken from black-forest-labs/FLUX.2-small-decoder. That repository ships no LICENSE file — the licence exists only as card metadata — so this mirror supplies an Apache-2.0 copy and a provenance README alongside it, as §4(a) requires of anyone redistributing the work.

Attribution

Created using Bonsai Image by Prism ML. Bonsai Image 4B is copyright 2026-present Prism ML, Inc., built from FLUX.2 [klein] 4B (copyright 2026 Black Forest Labs) with a text encoder built from Qwen3-4B (copyright 2024 Alibaba Cloud). All are available under the Apache 2.0 License; see each directory's NOTICE.md.

Apache-2.0 §6 grants no trademark rights. Naming the works here identifies their origin, as §4(c) requires, and is not a trademark licence.

Downloads last month

-

Downloads are not tracked for this model. How to track
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support