OsaurusAI/FLUX-3-Image-Klein-mflux-4bit

Non-commercial license

These weights are a quantized derivative of black-forest-labs/flux-3-image-klein and are distributed under the FLUX Kommunity License v1.0 (LICENSE.md). Any rights to use the model are granted directly by Black Forest Labs Inc. under that license. In short: non-commercial use, plus commercial use of generated images by Qualifying Users (under US$5M gross annualized revenue, including affiliates), subject to every term of the license, including its content-filtering and AI-disclosure conditions. For any other commercial use see https://bfl.ai/licensing. By downloading or using these files you agree to the LICENSE.

FLUX 3 Image Klein, mflux 4bit

An Osaurus 4bit mflux quantization of FLUX 3 Image Klein for Apple Silicon (5.0 GB, smallest). This is not an official Black Forest Labs release and is not endorsed, approved or validated by Black Forest Labs.

FLUX 3 Image Klein is the 16-step distilled model: a 5.6B-parameter transformer with a Qwen3-VL-4B text encoder. One bundle does text-to-image and reference editing (one to ten reference images). It also reads bounding boxes in the prompt and renders text inside images.

Variants

bundle	size	1024² generation (16 steps)	1024² edit (1 reference)
8bit	9.2 GB	44.8 s	89.7 s
6bit	7.1 GB	42.9 s	102.7 s
4bit	5.0 GB	40.8 s	82.2 s

Times are end to end, including the model load, on an Apple M5 Max (mflux).

Changes from the original (attribution notice: modified)

• Transformer and text-encoder linear layers re-saved with mflux and quantized to MLX affine 4-bit, group size 64. Norms, modulation, embeddings and biases are kept in bf16.
• The VAE is kept in bf16 except its 8 mid-block attention projections, which use the same 4-bit quantization.
• The text encoder holds the 32 Qwen3-VL language layers the model reads (it takes text features from layers 4 to 32); reference images are encoded by the VAE, so the vision tower is not included.
• No fine-tuning or other change.

Prompting

Write full scene captions: framing, subject, background, lighting. For edits, describe the whole target image with the change folded in. 16 steps, no guidance. Width and height must be multiples of 16.

Attribution and license

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

See LICENSE.md and NOTICE.md. The Qwen3-VL text encoder is under Apache-2.0. "FLUX" and "Black Forest Labs" are used only to describe the base model and its creator; no trademark license is granted or implied.

Quantization: Osaurus · mflux 0.23.0 (Flux3KleinEdit save path) · 4bit · eric@osaurus.ai