Good local image generation used to mean a big GPU. Black Forest Labs (BFL) is pushing that bar down: FLUX 3 Image Klein, released Oct 15, is a 5.6B-parameter open-weights model that BFL says runs on 8GB of VRAM. It was one of the most-discussed open models in our X sample this week, with about 2,500 posts.
What it does
According to BFL, Klein handles:
- Text-to-image
- Image editing, including what BFL calls pixel-perfect editing
- Multi-reference generation with up to 10 reference images
- Bounding boxes that control where elements go, or confine an edit to one area
- Native 4K generation, shared with the full FLUX 3 Image model
BFL's internal evaluations rank Klein first among open-weights models in text-to-image, multi-reference generation and editing, and it's the smallest model in that comparison. Against the previous FLUX.2 [klein] 9B, BFL reports gains of 91 Elo in text-to-image and 70 in editing, with 38% fewer parameters. These are BFL's own tests.
Hardware: what "8GB" really means
From BFL's announcement:
| Build | Generator weights |
|---|---|
| BF16 | 11.2 GB |
| FP8 | 5.6 GB |
| INT8 | runs on RTX GPUs with as little as 8GB VRAM (ComfyUI or Diffusers) |
BFL also ships NVFP4 builds, developed with NVIDIA. Speed, as measured by NVIDIA with the distilled model and INT8 build in ComfyUI: about 4 seconds per 1MP image on an RTX 5090 and under 9 seconds on an RTX 4070.
There's a catch. Those weight sizes don't include the text encoder (Qwen3-VL). Hugging Face's own Diffusers measurements, with everything loaded, peaked at:
| Setup | Peak after inference |
|---|---|
| NVFP4 (RTX 6000) | 15.3 GiB |
| FP8 (H100) | 16.27 GiB |
| BF16 (H100) | 20.72 GiB |
So on an 8GB card, use the INT8 build in ComfyUI, or in Diffusers load a quantized transformer, run the text encoder separately, and turn on CPU offloading. Hugging Face's post walks through all three.
License: free under $5M revenue
The weights use the FLUX Kommunity License. Per BFL, it allows free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use. Larger organizations need BFL's self-hosted licensing or the API.
If you're building a product, read the license itself and not just the summary. Licenses matter for owning your stack, as we cover in own your code and data.
Try it
Two checkpoints on Hugging Face:
| Checkpoint | What it is | Defaults |
|---|---|---|
black-forest-labs/flux-3-image-klein |
step-distilled, fast | 16 steps, no CFG |
black-forest-labs/flux-3-image-klein-base |
base, for fine-tuning | 50 steps, CFG |
In Python (Diffusers), from Hugging Face's post:
pip install -U diffusers transformers
from torch import bfloat16
from diffusers import ModularPipeline
pipe = ModularPipeline.from_pretrained("black-forest-labs/flux-3-image-klein")
pipe.load_components(dtype=bfloat16)
pipe.to("cuda")
pipe(prompt="a red fox", output="images")[0].save("fox.png")
That default loads in BF16, so it needs a bigger GPU. Use the low-memory options above on smaller cards.
No GPU? BFL's API offers flux-3-image-klein from $0.009 per 1K image, and there's a Hugging Face Space to test in the browser.
Prompting tip
Hugging Face notes that FLUX.3 "responds best to a full scene description with the subject, the setting, the light, the composition, and the style". So write a short paragraph, not a list of keywords.
Related
- Running text models locally as well? See Ollama for beginners and how much VRAM you need for a local LLM.
- Another open release this week: Flint vs Ornith local coding models.
