FLUX.2 [klein] 4B Base

The smallest FLUX.2 model: a 4B transformer conditioned on Qwen3-4B, using the same 32-channel FLUX.2 VAE and reference-image editing as FLUX.2 [dev]. The Base release is undistilled (no step or guidance distillation), Apache 2.0, and meant for fine-tuning.

weights: black-forest-labs/FLUX.2-klein-base-4B ↗
org: Black Forest Labs
modality: image
tasks: text-to-image · image-editing · multi-reference editing
license: Apache 2.0
released: 2026-01-15not verified
native output: 1024×1024not verified
total params: 7.98B
model.arch: flux2_klein_4b

Components

role	model	params	size	dtype	trained
Transformer	FLUX.2 [klein] 4B Base DiT black-forest-labs/FLUX.2-klein-base-4B/flux-2-klein-base-4b.safetensors ai-toolkit loads this single file in the original BFL layout into its own Flux2 module. The repo also has a Diffusers transformer/ folder with the same parameter count.	3.88B 3,875,544,576	7.75 GB	bf16	yes
Text encoder	Qwen3-4Bnot verified black-forest-labs/FLUX.2-klein-base-4B/text_encoder transformers.Qwen3ForCausalLM ai-toolkit loads Qwen/Qwen3-4B instead, which has the same parameter count (8,044,982,000 bytes across its shards).	4.02B 4,022,468,096	8.04 GB	bf16	no
Tokenizer	Qwen2 BPE (152k vocab) black-forest-labs/FLUX.2-klein-base-4B/tokenizer transformers.Qwen2TokenizerFast ai-toolkit loads the tokenizer from Qwen/Qwen3-4B.	—	—	—	—
VAE	FLUX.2 VAE (32 channels) black-forest-labs/FLUX.2-klein-base-4B/vae diffusers.AutoencoderKLFlux2 Same VAE as FLUX.2 [dev]. The klein repo only has it as a Diffusers folder, so ai-toolkit loads ai-toolkit/flux2_vae (ae.safetensors, fp32, 336,211,292 bytes, same parameter count) unless vae_path is set.	84.0M 84,046,372	168 MB	bf16+int64	no
total		7.98B	15.96 GB

Latent space

spatial

8×

channels

patch

2×2

autoencoder: FLUX.2 VAE
pixels per token: 16×16
notes: Same latent space as FLUX.2 [dev]: the VAE patchifies 2×2 and normalizes with BatchNorm, so the transformer sees 128-channel tokens at patch size 1, each covering 16×16 pixels.

input	latent (c×t×h×w)	tokens
512×512	32×1×64×64	1,024
1024×1024	32×1×128×128	4,096	ai-toolkit sample default
1344×768	32×1×96×168	4,032	landscape

Architecture

Blocks: 5 double-stream + 20 single-stream
Hidden size: 3072 (24 heads × 128)
FFN size: 9216 (mlp_ratio 3)
Text conditioning: Qwen3 hidden states from layers 9, 18 and 27, concatenated to 7680-d, 512 tokens. No pooled vector
Image conditioning: Reference images VAE-encoded and appended as tokens, each offset on the RoPE time axis
Guidance: No guidance embedding. Undistilled, sampled with real CFG
Objective: Rectified flow, resolution-dependent shift
Position: 4-axis RoPE (32, 32, 32, 32), theta 2000

In AI Toolkit

model.arch: flux2_klein_4b
UI label: FLUX.2-klein-base-4B (image)
model.name_or_path: black-forest-labs/FLUX.2-klein-base-4B
source: extensions_built_in/diffusion_models/flux2/flux2_klein_model.py
extensions_built_in/diffusion_models/flux2/flux2_model.py
extensions_built_in/diffusion_models/flux2/src/model.py
extensions_built_in/diffusion_models/flux2/src/pipeline.py
extensions_built_in/diffusion_models/flux2/src/sampling.py
toolkit/models/v2/vae/flux2_kl.py
toolkit/models/v2/text_encoders/qwen3.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: datasets.multi_control_paths, sample.multi_ctrl_imgs, model.low_vram, model.layer_offloading, model.qie.match_target_res

UI defaults

quantize / quantize_te: true / true
qtype: qfloat8
low_vram: true
train.unload_text_encoder: false
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
model_kwargs.match_target_res: false
network.conv: disabled (linear LoRA only)

Specifics

Code path: Flux2KleinModel subclasses the FLUX.2 [dev] model class; only the text encoder, VAE source and guidance handling differ.
Transformer source: flux-2-klein-base-4b.safetensors from name_or_path (a Hub repo or a local folder containing that file).
Text encoder source: Always Qwen/Qwen3-4B, whatever name_or_path is. Prompts go through the Qwen chat template with thinking disabled and are padded to 512 tokens. Never trained.
VAE source: ai-toolkit/flux2_vae (ae.safetensors) unless name_or_path is a folder with ae.safetensors or model.vae_path is set.
Resolution: Buckets snap to multiples of 16.
Reference images: Same as FLUX.2 [dev]: up to three control images per sample, datasets.multi_control_paths for training, each capped at 1 MP unless match_target_res is on.
Sampling: Not guidance-distilled, so samples run real CFG with the negative prompt at guidance_scale.
low_vram: Transformer and text encoder are loaded to the CPU and moved to the GPU when needed.
LoRA target: Flux2 (double_blocks and single_blocks)
Saving: LoRA keys use the diffusion_model. prefix and BFL module names (ComfyUI layout). LoKr saves in the new format by default. Full fine-tunes save one safetensors file.
Metadata base version: flux2_klein_4b

Example config

not verified

job: extensionconfig:  name: "my_flux2_klein_4b_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          cache_latents_to_disk: true          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 2000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "weighted"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16      model:        name_or_path: "black-forest-labs/FLUX.2-klein-base-4B"        arch: "flux2_klein_4b"        quantize: true        quantize_te: true        qtype: "qfloat8"        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 4        sample_steps: 30        prompts:          - "a bear building a log cabin in the snow covered mountains"