← Back to feed

Black Forest Labs released FLUX 3 Image Klein on October 15, officially. The weights had been on Hugging Face since October 13 — four repositories went up quietly under black-forest-labs/ before there was an announcement or technical report attached to them. The model was benchmarked and tested by third parties before BFL acknowledged it existed.

Klein is the open-weight, smaller member of the FLUX 3 Image family. The full FLUX 3 Image model is closed-API only. Klein is 5.6 billion parameters, trained with the same architecture and methods as its larger counterpart, and supports both image generation and image editing within a single checkpoint.

What runs where

Klein fits within an approximately 11.44 GiB device heap. That is within reach of gaming GPU configurations — most consumer cards sold in the past two years with 12 GB or more of VRAM can accommodate it. On hardware that can load the full model weights and run the Vulkan backend, the model runs without cloud dependency.

The practical floor for local inference is a GPU with 12 GB VRAM and support for a modern compute API. The model ships with a ComfyUI node for workflow integration and a prompt upsampler checkpoint (flux-3-image-klein-pup) for users who want to run automatic prompt expansion before generation.

Editing support is included without a separate checkpoint. Generation and inpainting/editing operate from the same model weights, which simplifies deployment compared to systems that require loading separate checkpoints for each mode.

Pricing and positioning

The API price for Klein is approximately 5x cheaper than full FLUX 3 Image on the same task. BFL has not published a per-image rate publicly, but third-party comparisons across major inference providers confirm the pricing differential is in that range.

For applications where image quality needs to be good rather than flagship-level — social content, product mock-ups, prototyping — Klein occupies a different cost tier than the full model. For local deployment, the API price is not the relevant number: inference runs at hardware cost.

Text rendering

Klein supports rendering Japanese text in generated images, which BFL highlighted in release materials. Historically, small image generation models have had poor multilingual text handling. Klein’s inclusion of Japanese support signals intentional training coverage for markets where CJK character rendering was previously a hard limitation of open models at this parameter count.

Competitive context

The open-weight image generation market has been dominated by FLUX 1 variants and Stable Diffusion checkpoints. FLUX 3 Image Klein moves the open-weight quality floor up to the same generational architecture as a commercially competitive closed model from the same lab, rather than a prior-generation open release.

The gap between closed and open image models has narrowed significantly over the past eighteen months. Klein’s release continues that pattern: architecture parity with the closed model, 5x lower API cost, and local deployment within consumer GPU memory limits.

The weights appeared before the announcement. That detail suggests BFL’s release process for open models is not fully controlled — or the team decided the community reaction to an unannounced drop was a useful signal before committing to a formal launch. Either way, the model had been benchmarked, tested, and discussed on technical forums for four days before BFL confirmed it was intentional.