FLUX 3 Image Klein: 4K AI Images on Your Own GPU
Black Forest Labs just did what the open-source image world has been waiting for. On October 15, 2026, the German AI lab released FLUX 3 Image Klein — a 5.6-billion-parameter image generation and editing model with open weights, native 4K output, and hardware requirements so modest it runs on a gaming GPU you might already own.
If you make images with AI — thumbnails, product shots, concept art, marketing creatives — this release changes the math. Until now, top-tier quality meant either paying per image through a closed API or running a 9B+ parameter model that demanded serious VRAM. Klein squeezes flagship FLUX 3 features (multi-reference generation, bounding-box layout control, image editing) into 5.6B parameters, and the INT8 build runs on as little as 8GB of VRAM. The API starts at less than a cent per image.
Here is what FLUX 3 Image Klein is, what the specs and pricing actually say, where it beats the competition, and where early testers say it still falls short.
What Is FLUX 3 Image Klein?
FLUX 3 Image Klein is a smaller, open-weights sibling of Black Forest Labs' flagship FLUX 3 Image model. One set of weights handles three jobs: generating images from text prompts, editing existing images, and multi-reference generation — feeding up to 10 source images to guide a single output (think: your product photo + a style reference + a background, composed into one shot).
The "Klein" name signals the point: this is FLUX 3 Image shrunk down. At 5.6B parameters it carries 38% fewer parameters than its predecessor, the FLUX.2 klein 9B variant — yet Black Forest Labs says it keeps the features that made FLUX popular with professionals: precise prompt following, strong text rendering, and now bounding-box control.
It is the second open-weights release in the FLUX 3 family, following FLUX 3 Action (the lab's 7B model for robot control). BFL says more releases are coming.
FLUX 3 Image Klein Specs
| Spec | FLUX 3 Image Klein |
|---|---|
| Parameters | 5.6B (image transformer; Qwen3-VL-4B-Instruct text encoder separate) |
| Weights | Open, on Hugging Face under FLUX Kommunity License |
| Output resolution | 768px square up to native 4K |
| Reference images | Up to 10 per request |
| Layout control | Bounding boxes (place elements, restrict edits to regions) |
| Variants | Klein Base (50 steps, CFG 3.5 — for fine-tuning) · Klein Distilled (16 steps, no guidance — fast inference) |
| Quantization | FP8, INT8, NVFP4 (NVIDIA co-developed) |
| Minimum VRAM | 8GB (INT8 build, consumer RTX GPUs) |
| Speed | ~4 sec per 1MP image on RTX 5090; <9 sec on RTX 4070 (vendor-measured, 16 steps) |
| Frameworks | ComfyUI, Diffusers, TensorRT-LLM, SGLang, vLLM-Omni (day-one support) |
| LoRA training | Diffusers or Ostris AI Toolkit on Klein Base |
The headline spec is the VRAM floor. Image models that can do serious editing work have typically wanted 12–24GB of VRAM, which prices out most hobbyists and freelancers. An 8GB requirement means a mid-range gaming card from the last few years can run it locally — no cloud bill, no queue, no content filter beyond your own judgment.
FLUX 3 Image Klein Pricing
The BFL API serves the distilled checkpoint as flux-3-image-klein. Text-to-image and image editing cost the same; you pay per image by resolution:
| Resolution | Price per image |
|---|---|
| 768px square | $0.008 |
| 1K | $0.009 |
| 1.5K | $0.011 |
| 2K | $0.014 |
| 4K | $0.036 |
For comparison, the full FLUX 3 Image model charges $0.048 (1K) and $0.607 (4K) for the same sizes — Klein is roughly 5x cheaper at 1K and nearly 17x cheaper at 4K. If you generate hundreds of images a month for a store, a channel, or client work, that gap is real money.
And if you self-host, the API price is irrelevant: the weights are downloadable, and your only cost is electricity.
The License: Free Commercial Use Has a Ceiling
The weights are released under the FLUX Kommunity License. The practical terms:
- Free for individuals and businesses under $5 million in annual revenue — including commercial use.
- Research and non-commercial use — free, no revenue check.
- Above $5M revenue — you need a paid self-hosted license or you use the API.
This is not MIT-style open source; it is "open weights with a commercial ceiling." For freelancers, small studios, YouTubers, and indie makers, the license is effectively free. For enterprises, it is a paid product. Know which side of the line you are on before building a business on it.
How to Use FLUX 3 Image Klein
Option 1 — The API (easiest). Same request format as FLUX 3 Image: prompt, images, aspect_ratio, resolution, plus bounding boxes. Switch the endpoint to flux-3-image-klein and existing FLUX 3 Image code runs on Klein. Prompt upsampling and pixel restoration (which preserves unedited regions exactly) are handled server-side.
Option 2 — ComfyUI (local, visual). BFL released ComfyUI weights and a Klein workflow on day one. Load the INT8 checkpoint, drop in the workflow, and you are generating on your own GPU.
Option 3 — Diffusers (local, code). Standard Hugging Face Diffusers pipeline; reference inference code and a Gradio demo for bounding-box control are on GitHub, plus a demo on Hugging Face Spaces if you want to try before downloading anything.
Option 4 — Fine-tune it. Train LoRAs on Klein Base with Diffusers or Ostris's AI Toolkit (arch: flux3); the trained LoRA also loads into the distilled checkpoint. NVIDIA's NeMo AutoModel ships training recipes for the released checkpoints.
FLUX 3 Image Klein vs the Competition
| Model | Parameters | Open weights | 1K image API price | 4K | Notes |
|---|---|---|---|---|---|
| FLUX 3 Image Klein | 5.6B | Yes | $0.009 | $0.036 | 8GB VRAM, 10 refs, bbox control |
| FLUX 3 Image (full) | Undisclosed | No | $0.048 | $0.607 | Flagship quality, closed |
| Nano Banana 2.1 | Undisclosed | No | ~5x Klein* | — | Leads Klein 7.20 vs 6.93 in one blind test |
| Qwen-Image 2.1 | Larger | Yes | Varies | — | BFL claims Klein beats it (vendor claim) |
| Ideogram V4.5 | Undisclosed | No | Varies | — | BFL claims Klein beats it (vendor claim) |
\Cost ratio from kie.ai's 200-test community relay, not official pricing.*
The honest read: on pure image quality, the one public third-party blind test (200 text-to-image and editing tasks, judged by Claude, Gemini, and GPT) scores Nano Banana 2.1 at 7.20 versus Klein's 6.93 — close, with Nano Banana ahead. But Klein costs about one-fifth as much per image in that test, ships open weights you can self-host and fine-tune, and runs on a $300 GPU. If you care about peak quality above all, the closed models still lead. If you care about cost, control, and deployment freedom, Klein is the most interesting release of the year.
For context on the wider field, see our 7 best free AI image generators and the Midjourney vs DALL-E vs Stable Diffusion breakdown.
Where It Falls Short (Honest Limitations)
- People and poses. Early hands-on testers report noticeably weak generation of people, odd poses, and viewpoint changes — apparently aggressive safety scrubbing of the training data is the suspected cause. If your work is portraits or character art, test carefully.
- Multi-reference compositing. Feeding it 10 reference images works, but testers say complex compositing across many references is shaky.
- Vendor benchmarks. BFL's claims of beating Qwen-Image 2.1 and Ideogram V4.5 are internal evaluations. Treat them as marketing until independent benchmarks land.
- Fine text rendering gets mixed reports — good in simple cases, weak in dense layouts.
- The license ceiling. $5M revenue cap means growing companies must re-license. Fine for most readers; a real consideration for agencies scaling up.
Who Should Use It?
- YouTubers and creators making thumbnails daily: the API cost is negligible and quality is strong for objects, scenes, and stylized art. (See our AI tools for YouTubers.)
- Indie makers and freelancers who want local generation with no per-image cost and no cloud dependency.
- Developers building image features into apps: the API is cheap, the request format matches FLUX 3 Image, and LoRA training is supported from day one.
- Small businesses under $5M revenue: free commercial use of genuinely good weights.
Skip it (for now) if your work is photorealistic people, or if you need the absolute best quality and don't care about cost — the closed frontier models still hold that crown.
Frequently Asked Questions
Is FLUX 3 Image Klein free?
The weights are free to download, and commercial use is free for individuals and businesses with under $5M in annual revenue under the FLUX Kommunity License. The hosted API costs $0.008–$0.036 per image depending on resolution. Above $5M revenue, commercial self-hosting needs a paid license.
What GPU do I need to run FLUX 3 Image Klein?
The INT8 quantized build runs on NVIDIA RTX GPUs with as little as 8GB of VRAM. BFL reports ~4 seconds per 1MP image on an RTX 5090 and under 9 seconds on an RTX 4070 using the distilled checkpoint at 16 steps.
What is the difference between Klein Base and Klein Distilled?
Klein Base is the undistilled checkpoint (50 steps, CFG 3.5) intended for fine-tuning and LoRA training. Klein Distilled runs 16 steps without guidance for fast inference — it is the checkpoint behind the BFL API. LoRAs trained on Base also load into Distilled.
Can FLUX 3 Image Klein edit images, not just generate them?
Yes. One set of weights handles text-to-image, single-image editing, and multi-reference generation with up to 10 reference images. Bounding boxes let you specify where elements go or restrict edits to a selected region, and the API's pixel restoration preserves unedited regions exactly.
How does it compare to Midjourney or DALL-E?
It is a different proposition: open weights you can self-host and fine-tune, versus closed subscription APIs. On quality, the one public blind test puts it slightly behind Nano Banana 2.1 (6.93 vs 7.20) at about one-fifth the cost. Against Midjourney specifically, Klein wins on price and control; Midjourney still leads on ease of use and community features.
Where can I download the weights?
Both checkpoints (Base and Distilled) are on Hugging Face under the FLUX Kommunity License, with reference inference code and a Gradio demo on GitHub, plus a try-it demo on Hugging Face Spaces.
Does it work with ComfyUI?
Yes — day-one support, with BFL-released ComfyUI weights and a Klein workflow. It also works with Diffusers, TensorRT-LLM, SGLang, and vLLM-Omni.
What are its weaknesses?
Early testers flag weak people/pose generation, shaky multi-reference compositing, and mixed fine-text rendering. Vendor benchmark claims (beating Qwen-Image 2.1, Ideogram V4.5) are unverified by third parties so far.
Reader reviews
Loading reviews…
