Black Forest Labs' FLUX 3 Klein: open weights, native 4K, runs on 8 GB of VRAM

Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, and a hands-on evaluation published on dev.to finds the 5.6-billion-parameter model has narrowed the gap with the strongest open rival to near parity: in a 100-image composite benchmark run conducted for the test, Klein scored 84.68 against 85.72 for Qwen Image 2.1, a difference of just over one point.

Open weights with a revenue cap

According to the dev.to write-up, Klein is distributed under the FLUX Kommunity License, which permits free commercial use for individuals and companies with annual revenue below $5 million. That makes the weights usable without an API key or waitlist, though the revenue threshold means larger businesses will need separate arrangements.

The spec sheet is compact for what the model does. Klein is described as the smallest model in its comparison class, with native 4K output at up to roughly 16 megapixels, text-to-image and image editing modes, multi-reference generation accepting up to ten input images, and bounding-box controls for layout. Weight files come in at 11.2 GB in BF16 and 5.6 GB in FP8, and INT8 builds run on 8 GB of VRAM.

Two checkpoints, one clear default

Black Forest Labs ships two variants: a base checkpoint that runs 50 sampling steps at CFG 3.5, and a step-distilled checkpoint that runs 16 steps without classifier-free guidance at CFG 1.0. Dropping guidance roughly halves sampling cost before the step reduction is even counted.

After 28 comparison tests, the tester's recommendation is to treat the distilled checkpoint as the default. The quality difference against the base model was smaller than the speed difference, and the distilled path rendered product typography more cleanly and held poses more steadily across seeds. The base model still earned its place for portraits, where heavier facial texture reads as more photographic, and for final deliverables where 50 steps of literal fidelity is worth the wait. Because both variants share the same latent space and VAE, seeds and resolutions transfer between them, which makes fair side-by-side comparison straightforward.

Seven ComfyUI workflows

The evaluation walks through seven ComfyUI workflows, four official graphs plus three derived ones. Among the practical guidance:

• Editing prompts should describe the finished frame rather than the change. Asking the model to "swap the earring" produced poor results, while describing the subject already wearing the new earring, with the same pose and lighting, worked.
• Multi-reference generation is strongest with one or two reference images. Accuracy falls off noticeably beyond four, and stacking the full ten means fighting for consistency.
• Bounding boxes, defined on a 0–1000 normalized grid, lean hard on placement without functioning as strict constraints, so slight drift is expected, especially on organic shapes. A useful technique is pinning anything that should not move by giving it identical source and target boxes — more reliable than asking in prose.

Strengths and shortcomings

Across the test runs, Klein performed solidly on detail fidelity for complex subjects such as full-body figures, hands and shoes; on local edits like object swaps, recolours and pose changes; and on character, outfit and scene fusion within the one-to-two reference range. Reported problems include consistency drift in multi-reference work, occasional limb failures on complex poses, and the soft nature of the bounding-box control. The tester notes that Qwen Image 2.1 shows the same failure modes, and that Klein's baseline is now high enough that these read as ordinary open-model limits rather than dealbreakers.

Running it locally

FP8 is recommended as the default for consumer cards, with BF16 reserved for quality-critical work on cards with 24 GB or more. Black Forest Labs reports roughly 4 seconds for a 1-megapixel image on an RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI — figures the tester flagged as vendor numbers rather than independently measured. Diffusers added day-one support through its ModularPipeline, though the prompt-upsampler component can detach from the block graph on newer transformers versions.

Why it matters

A one-point benchmark gap means open-weights image generation has effectively caught the frontier of its class, and the terms matter as much as the numbers: individuals and companies under the revenue cap can ship commercial work from a model that runs on an 8 GB card. For a generative image ecosystem that has watched capability concentrate behind APIs, the combination of competitive quality, permissive-enough licensing and consumer-hardware deployment is what makes Klein's release significant rather than incremental.