Black Forest Labs Ships FLUX 3 Image Klein, 5.6B Open Weights That Run on 8 GB GPUs

TL;DR

On October 15 Black Forest Labs put the first image model from its FLUX 3 family on Hugging Face: FLUX 3 Image Klein, a 5.6B-parameter transformer that does text-to-image, instruction editing and multi-reference generation with up to 10 reference images, at native 4K, in one checkpoint. BFL's internal evals put it 91 Elo ahead of its own 9B FLUX.2 klein on text-to-image and 70 Elo ahead on editing, with 38% fewer parameters. An INT8 build runs on 8 GB of VRAM through ComfyUI, and a 1 MP image takes under 9 seconds on an RTX 4070. The weights ship under a new FLUX Kommunity License that allows commercial use below $5M in annual revenue. The catch, per the people who ran it locally in the first three days: anatomy is shaky, and the 5.6B model wants a 9B prompt rewriter sitting in front of it.

What shipped

This is the second open-weights release in the FLUX 3 line, after the 7B robot-control model FLUX 3 Action, and the first one most readers can use. When BFL announced FLUX 3 in July as a single backbone trained across images, video, audio and action, open weights were a promise for later in the year, and this small checkpoint is the first payment on it.

Two checkpoints landed. The distilled flux-3-image-klein runs 16 steps with no guidance and is the one you generate with. The undistilled flux-3-image-klein-base runs 50 steps and exists for fine-tuning, LoRA training and research. Both share a 5.6B generator (11.2 GB in BF16, 5.6 GB in FP8) and a separate Apache 2.0 text encoder, Qwen3-VL-4B-Instruct. BFL also published FP8, INT8 and NVFP4 builds made with NVIDIA, a Hugging Face Space, and a reference repo whose code is Apache 2.0 and whose README recommends an 8 GB RTX 3070-class card use the INT8 files in ComfyUI.

Feature-wise, the headline is bounding boxes. You hand the model a layout on a 0 to 1000 grid, each box written as [y_min, x_min, y_max, x_max], and it places the element there or confines an edit to that region. The model card is careful to say boxes "do not constrain pixels to a hard mask," so treat it as a strong suggestion, not an inpainting mask. Multi-reference editing accepts up to 10 images, with the first reference setting the output size. Text rendering is supported and, per the same card, "may be inaccurate or subject to distortion."

The numbers BFL is claiming

All of the quality numbers here are BFL's internal evaluations, and the blog post gives Elo deltas rather than absolute scores. The company says Klein ranks first among open-weights models on text-to-image, multi-reference generation and editing, and rates above the closed Ideogram V4.5 and its own FLUX.2 [pro]. Against the 9B FLUX.2 klein the margin is +91 Elo on text-to-image and +70 on editing. It is also the smallest model in BFL's own comparison set, counting the generator only.

On latency BFL reports a median of 4.54 seconds end to end for a 1 MP text-to-image call on its API and 7.18 seconds for an edit, including queueing. Read the footnote before you repeat that: Klein was timed on BFL's own servers, and the competitors were timed from a client, with polling and network time included. The API lists the distilled model as flux-3-image-klein from $0.009 per 1K-resolution image.

What it costs you in seconds on a consumer card

The consumer numbers come from two places that agree. BFL's post says the distilled model in INT8 makes a 1 MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 in ComfyUI. The ComfyUI blog gives the before and after for the same cards: INT8 ConvRot quantization plus the step-distilled model is 4.3x to 9.3x faster than the BF16 50-step baseline, taking an RTX 4070 from 78 seconds to under 9 and an RTX 5070 from 71 seconds to under 8. At the other end of the budget, NVIDIA measured the distilled model at 0.55 seconds per 1024x1024 image on a GB300.

Step distillation is what makes the consumer numbers possible. The base model walks 50 denoising steps from noise to picture; the distilled student is trained to land on the same picture in 16, with no classifier-free guidance pass, so each step is cheaper and there are a third as many of them. Think of a chef who has cooked a dish a thousand times and now skips straight to the plating without tasting at every stage: faster, usually the same meal, and the one place it goes wrong is the fiddly bits. The reference repo's own figure is roughly 6x faster than base, and "the fiddly bits" turns out to be a literal description of the complaints below.

The 9B assistant the 5.6B model needs

The "5.6B" headline leaves out the helpers. The README states the model "ships with a prompt upsampler" and warns against using it without one; the model card says "a short prompt sent as typed gives generic output." The upsampler is a 9B vision-language model fine-tuned from Qwen3.5-9B, Apache 2.0, an 18.8 GB download, and it reserves 35% of GPU memory by default in the reference code. Add the 4B text encoder and the small model has about 13B parameters of helpers standing behind it. A 5.6B painter with a 9B art director rewriting every brief is a very 2026 org chart.

In practice the 8 GB path works because the upsampler runs separately and first. The README's small-GPU recipe is two commands: run flux3.pup to expand the prompt to a text file, then run the generator with --fp8 --no-pup on the expanded prompt. On a B200 at 1024x1024 with the upsampler off, peak memory is 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4, and the INT8 path that fits in 8 GB exists only in ComfyUI, not in the reference code. Budget around 52 GB of disk for the fast model plus its rewriter, about 107 GB for everything.

A license between Apache and "not for sale"

The FLUX.2 klein pair set the two poles: the 4B shipped Apache 2.0, and the 9B shipped under BFL's non-commercial license, so if you wanted the good one for a product you paid. FLUX 3 Image Klein splits the difference with the FLUX Kommunity License v1.0. Research and non-commercial use are allowed, and so is commercial use of the outputs for "Qualifying Users" with under US$5M in gross annualized revenue. Above that line you buy a self-hosted license or use the API. The code, the text encoder and the prompt upsampler are all Apache 2.0; only the generator weights carry the Kommunity terms.

For a solo creator or a small studio that is a real change from the 9B's terms: you can sell what Klein makes. For anyone whose revenue line might cross $5M, the license is an item on the legal checklist before this goes into a pipeline, and "we were under the cap when we started" is not a clause. Safety-wise, BFL says the checkpoints were dataset-filtered in pre-training, fine-tuned against harmful behaviors in post-training and red-teamed with the external partner Cinder, and claims "substantially fewer vulnerabilities than other popular open weights models."

What people found when they ran it

The Hacker News thread sat at 208 points and 57 comments when we read it, and the local reports split. The dominant complaint is anatomy. minimaxir wrote that a standard human test "has never failed on newer image models and it did fail; in addition to lower quality, the fingers were definitely not the correct proportions." vunderba reported the stock ComfyUI template producing "SD 1.5 levels of half-melted faces," and under the ComfyUI post the first comment reads "all eyes are distorted (tried base + destilled)." Several commenters connect this to the NSFW filtering in pre-training, which is a theory, not a finding; one of them pointed out the base checkpoint exists for LoRA training and predicted the obvious adapter.

The counter-reports are specific too. dragonwriter called the complaint "greatly exaggerated," said erroneous anatomy is "much less common than with Flux 2 Klein" when the whole stack runs, and noted some of the worst examples predate fixes to the ComfyUI templates, while conceding the distilled model "isn't great at generating photo-quality skin." jjcm said the editing side "has been really solid" and the weakness is fine text detail. The split lines up with the architecture: the distilled checkpoint is where the speed comes from and also where hands go wrong, and nobody in the thread is complaining about the edit path. One commenter observed that the last "3" out of this lineage was Stable Diffusion 3, and the thread took that about as kindly as you would expect.

Downloads have not waited for the verdict. The Hugging Face counters showed over 46,000 downloads of the distilled checkpoint and over 29,000 of the base three days after release. Day-one support covers ComfyUI, Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, with LoRA recipes in Ostris's AI Toolkit and NVIDIA NeMo AutoModel.

What to do with it

If you have an 8 to 12 GB card, pull the INT8 files from the ComfyUI repo, load the "FLUX 3 Image Klein" template, and test the editing and multi-reference paths first, because that is where both BFL's claims and the field reports agree it is strong. Run the upsampler; the authors say the model is tuned for its output and the generic results people get without it are not the model's fault. For people and hands, generate at the model's stable resolutions, try a second seed before you blame the weights, and watch the base checkpoint, because the first anatomy LoRAs will land there. If you ship commercially, read the Kommunity License before the model touches a product, and note the $5M line is gross revenue, not profit.

Key Takeaways

• FLUX 3 Image Klein is a 5.6B open-weights image model from Black Forest Labs, released October 15, doing text-to-image, editing and up to 10 reference images at native 4K in one checkpoint.
• BFL's internal evals put it +91 Elo on text-to-image and +70 on editing over its own 9B FLUX.2 klein, with 38% fewer parameters. No absolute scores were published.
• Consumer speed comes from step distillation plus INT8: an RTX 4070 drops from 78 seconds to under 9 per 1 MP image, and the INT8 build fits in 8 GB of VRAM via ComfyUI.
• The 5.6B figure excludes a required 9B prompt upsampler and a 4B text encoder; the fast model plus rewriter is about 52 GB on disk.
• The FLUX Kommunity License allows commercial use under $5M annual revenue, between the Apache 2.0 FLUX.2 klein 4B and the non-commercial 9B.
• Early local reports praise editing and flag hands and faces on the distilled checkpoint; the base model is where the fixes will come from.

Sources: Black Forest Labs: FLUX 3 Image Klein, Hugging Face model card, Klein Base model card, reference inference repo, ComfyUI blog: FLUX 3 Image Klein in ComfyUI, FLUX.2 klein 9B, FLUX.2 klein 4B, Hacker News discussion