Black Forest Labs Shrinks FLUX.2 Klein 9B Into a Lighter FP8 Build
Black Forest Labs ships an FP8 quantized checkpoint of its undistilled 9B klein base model, cutting VRAM needs while preserving the full training signal for fine-tuning.
Takeaways
- Black Forest Labs released an FP8 build of FLUX.2 klein 9B Base, fitting in ~29GB VRAM on an RTX 4090.
- Base means undistilled: full training signal preserved, ideal for LoRA training and custom fine-tuning pipelines.
- 9B rectified flow transformer paired with an 8B Qwen3 text encoder for stronger prompt adherence.
- Unified text-to-image, single-reference editing, and multi-reference generation with up to 5 reference images.
- Licensed under the FLUX Non-Commercial License; the sibling 4B klein ships under Apache 2.0.
- Part of the broader FLUX.2 klein family aimed at sub-second generation on consumer GPUs.
Black Forest Labs has published an FP8 checkpoint of FLUX.2 klein 9B Base. The release packages the undistilled foundation model in a lower-precision format for local inference, research, and supported adapter training. FP8 stores quantized tensors in 8-bit floating point, reducing their storage and memory footprint compared with BF16.
The release extends FLUX.2 klein, the compact branch of the FLUX.2 image-model family. Developers gain access to the adaptable Base weights with lower memory pressure, although the complete pipeline can still exceed the VRAM available on a 24GB GPU.
Base weights in eight-bit form
FLUX.2 klein 9B Base is a 9-billion-parameter rectified-flow transformer. Rectified-flow models learn a path from noise to an image, then follow that path during generation. The checkpoint supports text-to-image generation, single-image editing, and conditioning on as many as five reference images.
The 9B label refers to the image model. The full pipeline also uses an 8B Qwen3 text encoder to interpret prompts and rendered text, plus a variational autoencoder to translate between pixels and the model’s latent representation. Those additional components matter when calculating end-to-end memory use.
The FP8 repository contains a quantized copy of the Base model, while the full BF16 weights remain available separately. Quantization preserves the model’s task coverage but can produce small numerical differences from BF16 output.
Distillation sets the job
Klein checkpoints divide into distilled models optimized for short inference and Base models built for adaptation. Their practical differences are summarized below.
| Checkpoint | Sampling | License | Best fit |
|---|---|---|---|
| Distilled 4B | 4 steps | Apache 2.0 | Low-latency inference and permissive commercial deployment |
| Distilled 9B | 4 steps |