Black Forest Labs released FLUX 3 Image Klein on October 15, a 5.6-billion-parameter image model that generates, edits and composes from up to 10 reference images, with downloadable weights. BFL announcement The 5.6B count is the generator alone; the Qwen3-VL text encoder is separate, and Hugging Face’s safetensors metadata for the generator file lists about 5.55 billion BF16 parameters. Model card BFL says the INT8 build runs on GeForce RTX cards with as little as 8 GB of VRAM, and the same model is on its API from $0.008 an image. BFL announcement Release notes The FLUX Kommunity License grants commercial output rights only to users whose gross annualized revenue, counted with their affiliates, stays under $5 million. LICENSE.md
S5 Labs has not run the model. Every speed, memory and quality figure below is BFL’s unless attributed to someone else, and the leaderboard numbers were read on October 23 and will have moved.
Two checkpoints, two jobs
| Distilled | Base | |
|---|---|---|
| Steps | 16, no guidance | 50, CFG 3.5 |
| Precision shipped | BF16 (plus INT8 and NVFP4 in the ComfyUI repo) | Same sizes as distilled |
| Intended use | Generation and editing, self-hosted or via the API | Fine-tuning, LoRA training, research |
| Behind the BFL API | Yes | No |
Sources: distilled card, base card, ComfyUI repo.
The base card says it was trained without step or guidance distillation and that the distilled checkpoint is not the right starting point for training. Base card Train on Base, serve on Distilled. BFL’s docs give one worked example rather than a general recipe: a style LoRA trained on Base from 20 captioned public-domain paintings at about 1 megapixel, rank 32, alpha 32, learning rate 1e-4 with AdamW, 200 warmup steps and linear decay, 4,000 steps, about 4.5 hours on one GPU. BFL docs The docs then run that adapter on Distilled at LoRA scale 0.75 (0.5 to 1.0 works), 16 steps, CFG 1. BFL docs The docs also say a LoRA trained on Base loads into Distilled without any conversion, because the two share an architecture. The docs name AI Toolkit and Diffusers scripts as training tools, the blog adds NVIDIA NeMo AutoModel, and serving with a LoRA attached adds about 5% to generation time, per BFL’s GB300 figures. BFL announcement
What fits on an 8 GB card
Weight size and runtime memory are different numbers here, and BFL publishes only the first in detail. The ComfyUI repository lists the files:
| Component | BF16 | INT8 | NVFP4 / W4A8 |
|---|---|---|---|
| Transformer (distilled or base) | 11.1 GB | 5.8 GB (int8-convrot) | 5.3 GB (nvfp4) |
| Text encoder (Qwen3-VL-4B) | 8.9 GB | 4.9 GB | |
| VAE | 0.3 GB | ||
| Prompt upsampler (Qwen3.5-9B based) | 18.8 GB | 9.8 GB | 6.7 GB (w4a8) |
Source: ComfyUI repo.
The INT8 transformer plus the INT8 text encoder is 10.7 GB of weights before activations, so an 8 GB card cannot hold both at once. That sum is ours, not BFL’s, and it means the 8 GB claim depends on offloading one of them to system RAM. The Comfy README says the int8 files run on 8 GB GPUs such as the RTX 3070 and on cards without FP8 support (RTX 30 series and older), while the NVFP4 files need Blackwell. ComfyUI repo Neither the blog nor the README gives the resolution, batch size or offload settings behind the figure, so treat 8 GB as a vendor “as little as” line, not a tested minimum. BFL’s own docs pull the other way, saying Klein “fits on consumer GPUs, down to an RTX 3090,” a 24 GB card. BFL docs
The one place BFL reports measured memory is its reference code on a B200 at 1024x1024 with the upsampler off: Distilled peaks at 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4; Base at 31, 29 and 26 GB. Distilled card Base card Those are peaks on a datacenter GPU running reference code, so they say little about what a consumer setup with offloading needs. Turning the prompt upsampler on adds about 28 GB on an 80 GB GPU. Model card The docs give the same 28 GB without naming the GPU. Our local LLM machines guide covers the VRAM tiers this sorts into.
The prompt upsampler is not optional
The model card says not to use Klein without a prompt upsampler, because short prompts produce generic output. Model card The docs explain why: the model was trained on the captions the upsampler writes. BFL docs BFL’s upsampler is flux-3-image-klein-pup, a 9B model that the reference CLI loads by default. Diffusers does not call it for you, and the model card says any other LLM can take its place, given the right system prompt. Model card For edits, the same reference images should go to both the upsampler and the generator. BFL docs For a self-hoster this is the budget item: even the W4A8 build is 6.7 GB, more than the INT8 transformer, and the 8 GB claim says nothing about where it runs.
Speed, as BFL measured it
| Setup | Figure |
|---|---|
| Distilled INT8, ComfyUI, 1 MP, RTX 5090 | about 4 s |
| Distilled INT8, ComfyUI, 1 MP, RTX 4070 | under 9 s |
| Distilled NVFP4, GB300 | under 0.6 s |
| Base, 50 steps, 1024x1024, GB300 or DGX Station | 1.7 s |
| Base, 50 steps, 1024x1024, DGX Spark | up to 43.8 s |
| API median at 1 MP | 4.54 s text-to-image, 7.18 s editing |
Source: BFL announcement.
The 4070 figure is the one that matters for a mid-range desktop. The DGX Spark row, at up to 43.8 seconds, is why Base stays a training checkpoint.
The API: cheaper than self-hosting until the volume is large
The endpoint is POST https://api.bfl.ai/v1/flux-3-image-klein, with 1 to 16 steps (default 16), width and height from 64 to 6256 each, and a pixel product capped at 16,822,272, which is about 16.8 megapixels. API reference Text-to-image and editing cost the same, priced by resolution tier:
| Tier | Per image |
|---|---|
| 768sq | $0.008 |
| 1k | $0.009 |
| 1.5k | $0.011 |
| 2k | $0.014 |
| 4k | $0.036 |
Source: Release notes.
The blog’s “starting at $0.009 per 1K image” is the 1k tier, not the floor; 768sq is a tenth of a cent cheaper. BFL announcement Release notes Artificial Analysis lists Klein at $9.0 per 1,000 images, which matches the per-image reading. Artificial Analysis leaderboard At under a cent per megapixel image, self-hosting earns its keep when you need a LoRA the API does not serve, when images cannot leave the building, or when volume is large enough that a GPU pays for itself. The bounding-box inputs in the API examples appear to use a 0 to 1000 coordinate grid, going by values such as [0,0,1000,1000]; the docs do not define it. API reference
How good is it: the vendor’s numbers and an independent one
In BFL’s internal evaluation Klein ranks first in text-to-image, multi-reference generation and editing “ahead of larger open weights models,” and rates above the closed Ideogram V4.5 and FLUX.2 [pro]. Against its own FLUX.2 [klein] 9B it claims +91 Elo in text-to-image and +70 in editing with 38% fewer parameters. BFL announcement
Artificial Analysis runs a separate blind arena, and on October 23 it had FLUX 3 Image Klein at rank 21 (range 19 to 26), Elo 1043, confidence interval 1031 to 1055, on 5,184 votes. Qwen-Image-2.1 sits at rank 22 with 1042 and no API, and the two intervals overlap almost entirely, so read that as a tie. Fibo 2.0 is at 1039, BFL’s larger FLUX 3 Image at rank 13 with 1117 and $48 per 1,000 images, and FLUX.2 [dev] holds 1000 as the scale’s anchor rather than a measured score. Artificial Analysis leaderboard The vendor and independent results do not contradict each other, since the model sets and raters differ, but the board puts Klein level with Qwen-Image-2.1 rather than clear of it. The page marks open weights only with a Hugging Face link icon rather than a text label, and Elo drifts as votes come in, so check it before relying on the ranks.
Where Klein sits against the hosted editing models we have covered, Google’s Nano Banana 2.1 and OpenAI’s ChatGPT Images 2.5, is a separate question; the AA page we read covers text-to-image, not editing.
The license: free under $5M, with conditions
Section 1.e of the FLUX Kommunity License v1.0 defines a Qualifying User as “you, if at any time you together with your affiliates’ gross annualized revenue is less than US$5,000,000.” LICENSE.md The blog shortens that to “annual revenue under $5M,” but the text counts affiliates, says “gross annualized,” and leaves the measurement window unclear with “at any time.” The rest of the license:
- Qualifying Users may generate Outputs commercially (section 2.a). BFL claims no ownership of Outputs, and Outputs are not Derivatives.
- Access, use, distribution and creating Derivatives are otherwise limited to Non-Commercial Purposes (section 2.b), which excludes revenue-generating and production use. Derivatives include fine-tuned versions and other models that learn its patterns, behavior or outputs, so a merged LoRA checkpoint is one.
- Qualifying Users must run content filters or review Outputs before distribution, and disclose AI generation where the law requires (section 2.e).
- Redistribution needs an Attribution Notice; Derivatives also need a modification statement (section 3).
- The model may not be used to improve another AI model that performs a similar function (section 4).
- The license ends automatically on breach, ends if you sue over IP in the model or a Derivative, and BFL may terminate “at any time upon notice” (section 8).
Source: LICENSE.md.
A team can plan around a revenue line; it cannot plan around a termination-on-notice right. The diffusion weights carry this license, while the Qwen3-VL text encoder and the prompt upsampler are Apache-2.0. ComfyUI repo
Above $5 million, BFL’s blog points to a self-hosted license or the API. BFL announcement The licensing page lists four tiers, all covering Klein, with no prices: Builder (10K images a month, one domain, 10 licensed users, LoRA and fine-tuning allowed, internal and marketing use, not client work), Platform (100K images a month, one domain, 10 users, for SaaS and product teams serving end users), Professional (100K images a month, up to three domains, 10 users, first three clients included, per-client fees after) and Enterprise (custom). BFL licensing An agency over the line doing client work lands on Professional and pays per client; what that costs, BFL does not say publicly. Liquid’s LFM license, which we covered for LFM2.5-24B-A2B, draws a similar line at $10 million; BFL’s is half that and counts affiliates.
What to check before committing
- Measure VRAM on the card you have, at the resolution you need, with the encoder and upsampler placed where you intend to run them. No primary source gives a tested 8 GB configuration.
- If you train, train on Base and confirm the adapter behaves on Distilled at scale 0.75 before building a pipeline around the hand-off.
- Price the API against your volume at the tier you need before buying hardware.
- Have someone who reads contracts look at sections 1.e, 2.b and 8 together: affiliate revenue, the non-commercial limit on Derivatives, and termination on notice.
Our open-source AI models guide tracks where Klein’s license sits against Apache-2.0 alternatives. When BFL took FLUX 3 Video to general availability in August, the open-weight FLUX 3 Dev it mentioned had no date; Klein is the first FLUX 3 weights release we have covered.
Still unconfirmed
BFL’s blog names ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni as day-one runtimes; we have not checked the individual projects. BFL announcement The “convrot” in the INT8 filename is not explained anywhere we read. Self-hosted license prices are unpublished, and no one outside BFL has published a VRAM measurement at 8 GB.