FLUX 3 Image Klein: open weights, API access and license limits

Black Forest Labs released the open weights of FLUX 3 Image Klein on October 15, alongside a hosted API for the distilled version. For a creator or small technical team, the practical choice is between paying for a hosted image request and assembling a local inference setup. The two routes have different access steps, costs and licensing terms.

The big change

• What changed: FLUX 3 Image Klein's base and distilled checkpoints are available to download, and the distilled model is available through Black Forest Labs' API.
• Why it matters: Teams can inspect or adapt a checkpoint on their own infrastructure, but the download size, GPU memory requirement and commercial license each need a separate check.
• What to watch: Independent tests of output quality and resource use across complete local setups would show how broadly the vendor's hardware figures apply. BIG CHANGE has not run the model.

What the release includes

Klein's 5.6 billion parameter figure counts its image transformer. The official base model card says that a separate Qwen3-VL-4B-Instruct text encoder and an image autoencoder are part of the package. The model accepts a prompt and up to ten reference images and returns one image. The same model supports generation, editing and compositions that use references.

Black Forest Labs distinguishes the undistilled base checkpoint from the 16-step distilled checkpoint. The base is intended for fine-tuning, LoRA training and research; the distilled version is intended for generation, editing and self-hosting, and is the version behind the API. Downloading either repository's files requires accepting its Hugging Face access conditions and sharing contact information. The files are a model package, not a finished image application.

Local inference also needs an implementation that supports the checkpoint and its quantized format, if one is selected. The model card points to Black Forest Labs' reference code and Diffusers. Its reference command loads a separate safety classifier and, by default, a prompt upsampler. The latter converts a short request into the full scene caption the model expects. The model card warns that sending a short prompt unchanged tends to produce generic output. This is a documented setup requirement, not a result BIG CHANGE reproduced.

Download size is not GPU memory

Black Forest Labs' release post lists 11.2 GB for the generator weights in BF16 and 5.6 GB in FP8. Those figures describe generator weights. They do not include every component or the memory used while generating an image. The base model card reports peak GPU memory of 31 GB in BF16, 29 GB in FP8 and 26 GB in NVFP4 for its reference code on a B200 at 1024 × 1024, with prompt upsampling off. It says the prompt upsampler adds about 28 GB on an 80 GB GPU.

The company separately says an INT8 build can run on an NVIDIA RTX GPU with as little as 8 GB of VRAM through ComfyUI or Diffusers. NVIDIA's reported 1024 × 1024 measurements for the distilled INT8 build are about four seconds on an RTX 5090 and under nine seconds on an RTX 4070 through ComfyUI, at 16 steps. Those are particular optimized builds and measured conditions. They are not the memory or speed of the BF16 reference setup, and BIG CHANGE has not independently measured them.

The API route and its price

The FLUX 3 Image Klein API reference specifies POST https://api.bfl.ai/v1/flux-3-image-klein with an x-key API credential. A request needs a text prompt. It can also include one to ten reference images as URLs or base64 data; the first reference determines the default output size for an edit. Callers can set an aspect ratio or resolution tier, or provide width and height together. The default output file format is JPEG, with PNG available.

The submission response gives a task ID and polling_url, rather than the finished image. The Get Result endpoint accepts that ID and reports task status and result. This means an application must handle an outstanding job and possible error or moderation result. The API documentation does not establish a fixed completion time for every request.

The release post says hosted pricing starts at $0.009 per 1K image. Actual charges depend on the request; the submission response includes a cost field in credits. Local deployment has no per-image API charge in this route, but needs hardware, storage and an inference setup. Neither path's total cost follows from the price or weight size alone.

What the weights license permits

The downloadable checkpoints use FLUX Kommunity License v1.0. It permits non-commercial uses under its terms. An individual or entity whose combined gross annualized revenue with affiliates is below US$5 million qualifies for a limited exception to generate and use commercial outputs, including production use for that purpose. That exception includes content filtering or review obligations and applicable AI disclosure requirements. It does not grant unrestricted commercial fine-tuning, model distribution or use outside the license's stated scope.

Black Forest Labs directs other commercial deployments to its separate self-hosted licenses or API. The model card lists further restricted uses, including surveillance, biometric processing and using outputs to improve a similar AI model. Teams considering local use should read the full license alongside their intended workflow; the Apache-2.0 license on the separate text encoder does not replace the terms for Klein's weights.

The release gives teams another way to work with FLUX 3 images, while leaving two decisions to verify for each deployment: whether a complete inference stack fits the available machine and whether the intended use fits the applicable license. BIG CHANGE reviewed the public documentation and did not download the weights, install an inference stack, submit an API request or compare generated images.

Sources & further reading

• Black Forest Labs, “FLUX 3 Image Klein”, October 15, 2026. Establishes the release, advertised model variants, hardware measurements, API availability and starting price. Quality and speed results are vendor-reported; BIG CHANGE did not reproduce them.
• Black Forest Labs, FLUX 3 Image Klein Base model card, read October 18, 2026. Explains the package components, base versus distilled roles, download access, prompt upsampling, reference-code memory figures, limitations and license summary. Its documented steps were not run by BIG CHANGE.
• FLUX Kommunity License v1.0, read October 18, 2026. Governs the weights and defines the qualifying-user commercial-output exception and its conditions; readers should use the full text for their own use case.
• BFL FLUX 3 Image Klein API reference and Get Result reference, read October 18, 2026. Specify the request fields, asynchronous task ID, polling and response schema. BIG CHANGE made no API call.