Bonsai Image 4B (GGUF)

Prism ML's Bonsai Image 4B, built on FLUX.2 [klein] 4B, packaged for stable-diffusion.cpp for making pictures on high-memory Adreno phones; current TokForge releases no longer offer it.

Published by TokForge for on-device inference.


What it is	Image generator
Format	stable-diffusion.cpp bundle: TQ2G128 GGUF diffusion transformer, Qwen3-4B Q4_K_M GGUF text encoder, FLUX.2 VAE (safetensors)
Runs on	Android phones with a Qualcomm Adreno GPU (stable-diffusion.cpp, OpenCL); other phones get a Preview route
Phone class	Adreno phones with 12 GB RAM or more (beta); 16 GB or more for the Quality Beta route
Download	4.03 GB (8 files)
Base model	prism-ml/bonsai-image-ternary-4B-gemlite-2bit
License	Apache-2.0 (Bonsai Image, FLUX.2 [klein] 4B and Qwen3-4B); attribution notices in `NOTICE.md`

TokForge

Website: https://tokforge.ai
Discord: https://discord.gg/Acv3CBtfVm
Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
iOS TestFlight: https://testflight.apple.com/join/jnufjzRr

Runs on-device in the TokForge app.

Use it in TokForge

Install TokForge from Google Play.
Current TokForge releases no longer offer this bundle in the model catalog, so it cannot be downloaded from the app. A copy that is already installed keeps working.
With it installed, ask for a picture in a chat, for example generate an image of a red car on a city street. TokForge unloads and reloads the chat model when needed on phones with less memory.

Collections: TokForge GGUF Models, TokForge Image Models (Android).

Files

File	Size	What it is
`bonsai_klein_official_tq2g128.gguf`	1.37 GB	Bonsai Image 4B diffusion transformer, TQ2G128 GGUF for stable-diffusion.cpp
`flux2-vae.safetensors`	168 MB	FLUX.2 [klein] VAE
`manifest.json`	1.68 KB	Component sources, sha256, recommended settings and route policy
`Qwen3-4B-Q4_K_M.gguf`	2.50 GB	Qwen3-4B text encoder, GGUF Q4_K_M
`samples/adreno_quality_beta_contact_sheet.png`	1.43 MB	Sample pictures from the Adreno Quality Beta route
`LICENSE`	10.2 KB	Apache-2.0 license text
`NOTICE.md`	623 B	Attribution notices (Prism ML, Black Forest Labs, Alibaba Cloud)
`upload_commands.sh`	580 B	Helper script used to upload this repo (not needed to use it)

File	Bytes	SHA-256
`bonsai_klein_official_tq2g128.gguf`	1,367,729,664	`2a6da84102513a3a14a955edeb84ef0e347b51f9f13cc539c939ea78410a3eb7`
`Qwen3-4B-Q4_K_M.gguf`	2,497,280,256	`7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5`
`flux2-vae.safetensors`	168,120,878	`ca70d2202afe6415bdbcb8793ba8cd99fd159cfe6192381504d6c4d3036e0f04`

The sample contact sheet in samples/adreno_quality_beta_contact_sheet.png shows Adreno Quality Beta outputs for car, tree, restaurant, beach, house, and person prompts.

About the model

TokForge-packaged Bonsai Image 4B / FLUX.2 Klein bundle for on-device Android image generation:

bonsai_klein_official_tq2g128.gguf: TokForge/sd.cpp GGUF conversion of the Bonsai Image 4B diffusion transformer.
Qwen3-4B-Q4_K_M.gguf: Qwen3-4B text encoder in GGUF Q4_K_M format.
flux2-vae.safetensors: FLUX.2 Klein VAE, renamed for TokForge bundle discovery.

Attribution: Bonsai Image is by Prism ML. FLUX.2 Klein is by Black Forest Labs, and the text encoder is based on Qwen3-4B by Qwen/Alibaba Cloud. This repository changes the deployment packaging and runtime formats only.

Where it runs

This is a guarded beta image-generation bundle, not a universal default for every Android phone.

Device class	Route	Status
16-24 GB Qualcomm/Adreno	512x512, OpenCL, 2 steps	Quality Beta / Lab
12-16 GB Qualcomm/Adreno	512x512, OpenCL, 2-3 steps	Quality Beta / Lab after device proof
CPU, Tensor, Exynos, Mali, Xclipse	sub-512 or CPU preview	Preview only

Quality Beta means coherent 512x512 output, repeatable, with no flat or corrupt image.
Preview is used for sub-512, CPU, low-memory, or visibly soft/warped outputs.
Lab is used for OpenCL/server paths that require explicit opt-in.

Render settings

The Bonsai route expects the published low-guidance FLUX.2 Klein/Bonsai-style recipe:

Setting	Value
Sampler	Euler
Scheduler	simple / FlowMatch-style
CFG scale	1.0
Distilled guidance	1.0
Flow shift	3.0
Default canvas	512x512
Default quality steps	2 on supported Adreno OpenCL

Higher step counts are not automatically better; 1-step output is too smeared for Quality Beta.

How it was made

Converted the diffusion transformer from prism-ml/bonsai-image-ternary-4B-gemlite-2bit (via prism-ml/bonsai-image-ternary-4B-unpacked) to a TQ2G128 GGUF with stable-diffusion.cpp.
Took Qwen3-4B-Q4_K_M.gguf unchanged from Qwen/Qwen3-4B-GGUF (byte-identical).
Took the FLUX.2 [klein] 4B VAE unchanged (byte-identical) and renamed it flux2-vae.safetensors.

manifest.json records each component's source, sha256 and the recommended settings.

Limitations

This is a runtime bundle, not a standard Diffusers training checkpoint.
OpenCL acceleration is guarded and targeted at high-memory Qualcomm/Adreno devices.
Non-Adreno GPU paths are not release-green.
Low-memory CPU routes are Preview only.
The output quality gate is visual/coherence-first; blurry or flat routes are not labelled Quality Beta.

License

Every part of this bundle is under the Apache-2.0 license: Bonsai Image by Prism ML, FLUX.2 [klein] 4B by Black Forest Labs (the VAE) and Qwen3-4B by Alibaba Cloud (the text encoder). The license text is in LICENSE and the attribution notices are in NOTICE.md. Prism ML asks for attribution such as: "Created using Bonsai Image by Prism ML."

Credits

Bonsai Image ternary 4B by Prism ML (Hugging Face). Used under the Apache-2.0 license. Created using Bonsai Image by Prism ML.
Bonsai Image ternary 4B unpacked by Prism ML (Hugging Face). Used under the Apache-2.0 license.
FLUX.2 [klein] 4B by Black Forest Labs (Hugging Face). Used under the Apache-2.0 license. The VAE file is byte-identical to its VAE.
Qwen3-4B GGUF (text encoder) by the Qwen team (Alibaba Cloud) (Hugging Face). Used under the Apache-2.0 license. Qwen3-4B-Q4_K_M.gguf is byte-identical to it.
Converted to GGUF with stable-diffusion.cpp by leejet (MIT).