license: other
license_name: civitai-model-license
license_link: https://civitai.red/models/2373678/uncensored-qwen3-4b-z-image-and-flux4b
base_model: black-forest-labs/FLUX.2-klein-4B
tags:
- openvino
- flux2
- flux2-klein
- qwen3
- text-encoder
- int4
- not-for-all-audiences
Uncensored Qwen3 4B · FLUX.2 klein 4B text encoder · OpenVINO INT4
OpenVINO INT4 conversion of Uncensored Qwen3 4B (Z-Image & Flux4B) by Felldude, a finetune of the
Qwen3-4B text encoder, made to replace the text encoder of
FLUX.2 [klein] 4B.
- Original model: Civitai,
version "FP32", fileqwen3_4b_human_corpus_fp32.safetensors - All credit for the finetune goes to its author. This repository only changes the format.
This is not a general-purpose LLM. Only the 27 Qwen3 layers FLUX.2 klein reads are kept, and the model
outputs klein's prompt_embeds (hidden states of layers 9, 18 and 27, concatenated: 7680 values per token). It
works only as the text encoder of FLUX.2 klein 4B (not Z-Image, which reads other layers).
Usage
Made for the FLUX.2 klein 4B OpenVINO web UI: download it from the Download model tab and select
it as the text encoder. It can be combined with any FLUX.2 klein 4B transformer (the original or a finetune). The
files are in text_encoder/, with the same inputs and outputs as the text encoder of
rupeshs/flux2-klein-4b-int4-ov, so it can also replace
that folder.
How it was converted
The finetuned weights were quantized into the OpenVINO graph of the text encoder of rupeshs' export, with the same
precision for every weight (INT4 asymmetric with groups of 128 for most layers, INT8 per channel for the rest, as
NNCF does): same size (~2.1 GB) and speed. Checked against the FP32 finetune run in PyTorch: cosine similarity of
the output embeddings 0.98–0.99 for layers 9/18/27 (the difference is the INT4 rounding; the original text encoder
gives 0.82–0.97 against the same reference). Conversion script: tools/convert_text_encoder.py in the web UI
repository.
Credits and license
- Finetune: Felldude (Uncensored Qwen3 4B on Civitai);
the permissions set by the author on Civitai apply. - Base text encoder: Qwen3-4B by the Qwen team (Apache 2.0), as used by
FLUX.2 [klein] 4B (Apache 2.0). - OpenVINO graph: rupeshs/flux2-klein-4b-int4-ov (Apache 2.0).