Black Forest Labs Releases FLUX 3 Image Klein, a 5.6B Open-Weight Image Model

The model and usable artifacts arrived on October 15

Black Forest Labs announced FLUX 3 Image Klein on October 15, 2026 and made it available through the BFL API and two official Hugging Face repositories. The production-oriented identifier is black-forest-labs/flux-3-image-klein; black-forest-labs/flux-3-image-klein-base is the undistilled training and research variant. At review time, both pages and file manifests were public, but downloading required a signed-in user to accept conditions and share contact information.

The distilled API model is named flux-3-image-klein. BFL says it includes prompt upsampling and pixel restoration for edits. The public reference code, examples and runtime recipes are also available, so builders can inspect the serving path before accepting the weight gate. “Open weights” here means downloadable under BFL’s terms; it does not mean unrestricted access or an OSI open-source license.

One 5.6B image transformer handles generation and editing

Klein supports text-to-image, single- and multi-reference editing, and composition from as many as ten reference images. It also accepts bounding boxes for placing elements or limiting an edit to a region. BFL advertises output up to 4K. These are documented modes and limits, not an AccessAllGPT finding about instruction following, identity preservation, typography or edit fidelity.

The 5.6B figure describes the image transformer only. The model card separately names Qwen3-VL-4B-Instruct as the text encoder, while BFL’s reference stack uses a 9B prompt upsampler derived from Qwen3.5-9B. Teams comparing parameter counts or memory therefore should inventory the complete pipeline rather than treating 5.6B as the total resident system.

Distilled and base checkpoints serve different decisions

The distilled checkpoint generates in 16 steps without classifier-free guidance and is the model behind the BFL API. The base checkpoint uses 50 steps with guidance 3.5 and is intended for fine-tuning, LoRA training and research. BFL says the distilled path is about six times faster in its reference setup; whether that trade preserves the details a product needs must be tested on the product’s prompts and edits.

Prompt expansion is part of the intended pipeline, not an optional cosmetic feature. The official code warns that short prompts sent directly produce noticeably weaker results and enables its prompt upsampler by default. A fair evaluation should pin the upsampler, its system prompt and the image model together, while also recording the latency, memory, moderation and privacy effects of that extra model call.

Consumer-GPU claims need full-pipeline memory checks

BFL says the image-generator weights occupy 11.2 GB in BF16 and 5.6 GB in FP8, and that an INT8 ComfyUI build can run on NVIDIA GPUs with 8 GB of VRAM. Its launch post reports roughly four seconds for a one-megapixel distilled image on an RTX 5090 and 8.9 seconds on an RTX 4070. NVIDIA measured the cited GPU times, using one image per run and the fastest listed configuration; AccessAllGPT did not reproduce them.

Those compact weight figures are not end-to-end peak memory. The reference repository reports roughly 25–31 GB of peak GPU memory on a B200 for its BF16, FP8 and NVFP4 CLI paths without the upsampler, and says the upsampler reserves additional GPU memory by default. It describes a 52 GB download for the fast model plus upsampler and 107 GB for all documented weights. INT8, FP8 and NVFP4 also have different runtime and hardware requirements, so test the exact package rather than extrapolating from one weight number.

The API starts at $0.009 for a 1K image

BFL’s launch post lists a starting price of $0.009 per 1K image. In this context, “1K” describes image resolution; it does not mean one thousand images. The live pricing documentation should be the purchasing source of truth because output size, operation and future price changes can alter the bill. The announcement does not establish a workload-wide cost after prompt expansion, retries, storage, moderation and rejected outputs.

For a real comparison, hold prompts, source images, output size, seed policy, quality review and retry rules constant across the BFL API and self-hosted candidates. Measure cost per accepted asset, edit success, unmodified-region fidelity, latency percentiles, queue failures and reviewer time. A low list price can still lose when a workflow needs repeated generations or manual repair.

The weight license has a commercial revenue boundary

The model weights use the FLUX Kommunity License, not Apache 2.0. BFL summarizes it as allowing free commercial use for individuals and organizations below $5 million in annual revenue, plus research and non-commercial use. Larger organizations are directed to BFL’s self-hosted licensing or API. Teams should review the actual accepted license text and definitions with counsel; this article is not legal advice.

The reference inference code is Apache 2.0, and BFL says the separate prompt upsampler is Apache 2.0. Those licenses do not replace the weight terms. Record the model revision, accepted license version, organization revenue analysis, deployment entity, derivatives and distribution plan before fine-tuning or embedding weights in a product.

Vendor evaluations do not settle image quality

BFL reports that Klein leads its internal comparisons for text-to-image, multi-reference generation and editing, including gains of 91 Elo for text-to-image and 70 Elo for editing over FLUX.2 Klein 9B. It also reports median BFL API latency of 4.54 seconds for text-to-image and 7.18 seconds for editing at one megapixel. These are vendor-selected evaluations; competitors were not all measured through identical infrastructure, and AccessAllGPT did not inspect prompts, raters or raw votes.

Evaluate the jobs the product actually performs: brand styles, people and products, non-English text, multiple references, layout boxes, local edits, unchanged-region preservation and refusal behavior. Blind review should separate prompt fidelity, visual quality and edit correctness. Preserve prompts, references, seeds, model and upsampler revisions, raw outputs and failure counts so a later checkpoint or API revision can be compared honestly.

What builders should do next

Use the API for a bounded first trial when its data terms and service geography fit: pin flux-3-image-klein, fix one or two output sizes, include representative generation and editing tasks, and cap retries. Compare it with the exact self-hosted variant the available GPU can run. Do not compare BFL’s best API output with an under-provisioned local path that omits the required prompt upsampler.

Advance to self-hosting only after accepting the gate and license, scanning and pinning artifacts, validating complete-pipeline memory, enabling an appropriate safety policy, and measuring quality and cost. Wait when the license boundary, training-data diligence, moderation design or hardware path is unresolved. Reject the model if local tests fail quality, edit fidelity, legal, privacy, safety, latency or cost gates. Availability is the start of the decision, not the decision itself.

Disclosures

AccessAllGPT received no weights, API access, credits, hardware, images, benchmark prompts, votes, logs, private briefing, review, payment or compensation for this article. Black Forest Labs, Hugging Face, NVIDIA and the named runtime projects did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank FLUX 3 Image Klein. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Black Forest Labs, Hugging Face or NVIDIA. Publication-wide relationships are listed on the disclosures page.