On October 15, Black Forest Labs released FLUX 3 Image Klein. It is a 5.6B-parameter open-weight model that generates images from text, edits images, and accepts up to 10 reference images at once.
The most striking part is not the parameter count or the benchmarks. It is where it runs. The company says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM. Given that image models in the 12B to 20B range have recently demanded 24GB, the direction has clearly changed.
This piece covers what the model can do, the conditions behind those numbers, and how far to trust it.
What it can do
From the official announcement, the capabilities line up like this.
FLUX 3 Image Klein (Black Forest Labs, 2026-10-15)
Base : 5.6B-parameter open-weight image generator
Features : text-to-image, image editing, multi-reference
generation with up to 10 reference images
Resolution : native 4K generation
Control : bounding boxes that place elements or confine
edits to a defined area (pixel-perfect editing)
Checkpoints: a base model for fine-tuning, plus a step-distilled
model for low-latency generation
Builds : FP8, INT8 and NVFP4 (with NVIDIA support)
License : FLUX Kommunity License
Day one : ComfyUI, Hugging Face Diffusers, TensorRT-LLM,
SGLang Diffusion, vLLM-Omni
Training : Ostris AI Toolkit, NVIDIA NeMo AutoModel (LoRA)
Pay attention to bounding-box control. It lets you say "place this element here" or "confine the edit to this region" as coordinates. One of the hardest problems in generative editing is not touching what you did not want to change. The API documentation mentions a pixel restoration feature that preserves unedited regions exactly.
Memory: what fits where
Memory is the heart of this release. Here are the figures from the announcement.
| Item | Size | Where it runs |
|---|---|---|
| Generator weights (BF16) | 11.2GB | Plenty of VRAM recommended |
| Generator weights (FP8) | 5.6GB | Mid-tier VRAM |
| INT8 build | as little as 8GB VRAM | RTX GeForce GPUs (ComfyUI/Diffusers) |
| NVFP4 (GB300) | — | Server class, under 0.6s for 1MP |
One condition is easy to miss. Every parameter count here (5.6B, and the 7B and 12.9B comparisons) refers to the image generator only. Text encoders are excluded. Read a comparison table without that condition and the real-world feel will not match.
Speed: read it with the conditions attached
Generation time is published concretely too.
[Time for a 1MP image, distilled model at 16 steps]
RTX 5090 : about 4 seconds
RTX 4070 : under 9 seconds
GB300(NVFP4): under 0.6 seconds
Conditions
- Measured by NVIDIA; RTX results are in ComfyUI
- Each GPU uses its quickest build
- This is time on the GPU itself. API latency adds queueing,
network and prompt upsampling on top
- Serving a LoRA adds roughly 5% to generation time
"Time on the GPU itself" is the key qualifier. The time until a finished image reaches you includes queueing and network, so it is longer.
Quality: first in internal evals, with strings attached
Black Forest Labs says that in its internal evaluations the model ranks first in text-to-image, multi-reference generation and image editing — ahead of larger open-weight models, and above the closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro].
[Where it sits in the internal comparison]
- At 5.6B, the smallest open model in the group
(Qwen-Image 2.1 at 7B, Krea 2 at 12.9B)
- Against FLUX.2 [klein] 9B:
+91 Elo in text-to-image
+70 Elo in image editing
38% fewer parameters
- Krea 2 does not support image editing, so it is excluded
from the editing comparison
[Caution]
- These are the vendor's internal evaluations
- Parameter counts are for the image generator; text encoders excluded
Speed and quality are shown together. On the BFL API, the model delivers lower latency than Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5 and Krea 2, and the highest rating at its latency and above. In other words, it claims "faster, or the same speed but better," traced as a Pareto curve. API pricing starts at $0.009 per 1K image.
8GB does not solve everything
From here, the practical caveats — and the announcement states them itself.
What "supported on 8GB of VRAM" precisely means
- Runs with as little as 8GB, for the INT8 build only
- Memory management and CPU offloading may be necessary
- It does not guarantee that every high-resolution or
multi-reference task fits in 8GB
That produces a practical split.
| Task | Realistic on 8GB? |
|---|---|
| 1MP generation (distilled, INT8) | Yes, with some slowdown from offloading |
| Editing with a single reference | Usually yes |
| Multiple reference images | Heavy regardless of memory; may struggle on some cards |
| Native 4K | Likely beyond 8GB |
One more thing to check: the license. The FLUX Kommunity License allows free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use. Larger organizations need self-hosted licensing or the API. "Open weights" does not mean "free with no strings."
On safety, the company notes dataset filters during pre-training, targeted fine-tuning to suppress harmful behavior during post-training, and internal red-teaming plus evaluation by external partner Cinder ahead of release.
Conclusion: judge local image models on four axes
The briefing's action, turned into an experiment:
[A mini experiment for a local image model]
1. Choose one editing task
(e.g. swap a product photo's background, fix portrait tones)
2. Run the candidates on the same input
[ ] same resolution, same step count, same quantization (INT8)
[ ] same prompt and bounding boxes
3. Record four axes
[ ] Quality : how well unedited regions are preserved
[ ] Latency : request to finished image
[ ] Memory : peak VRAM (with or without offloading)
[ ] License : commercial scope and revenue conditions
4. Reproduce it in a ComfyUI or Diffusers workflow
One line captures it.
Image-model competition is extending from
"how well does it draw" to "how little hardware can it run on."
FLUX 3 Image Klein is an important milestone in that direction. Still, "runs on 8GB" should be read as "the floor came down," not "8GB is all you need." Before shipping, measure the peak memory and latency of your own workload. For related threads, see Krea 3: Fast Image Editing and Meta Muse Image Editing.
Further reading
- Krea 3: What Fast Image Editing Really Means
- Meta Muse Image Editing: What Changed
- Generative AI Content Basics: Text, Image and Video
References
- FLUX 3 Image Klein announcement (Black Forest Labs, Oct 15, 2026)
- Black Forest Labs pricing
- ComfyUI official site
Go deeper with a course
Image generation and editing click fastest when you change prompts and compare results, over and over. Treat writing, images and video as one flow, and you start to see which tool fits where.