Black Forest Labs ships FLUX 3 Image Klein, a 5.6B model running 4K generation on 8GB GPUs

A smaller FLUX aimed at consumer hardware

Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, according to a toolpilotpro guide syndicated on dev.to. The model packs 5.6B parameters, ships with open weights and supports native 4K output, while keeping hardware requirements low enough for a mid-range gaming GPU. The pitch is straightforward: capable image generation on machines people already have, rather than rented data-center compute.

One model, several jobs

A single set of weights covers three tasks: text-to-image generation, image editing, and multi-reference generation in which up to 10 source images guide one output. It also supports bounding-box layout control, so users can place an element by drawing a box or confine an edit to a selected region.

Two checkpoints ship. Klein Base runs 50 steps and is intended for fine-tuning and LoRA training, while Klein Distilled runs 16 steps and is the quicker variant served behind the API.

Runs in 8GB of VRAM

The INT8 build runs on NVIDIA RTX cards with as little as 8GB of VRAM. Per the guide, Black Forest Labs measures about 4 seconds per 1-megapixel image on an RTX 5090, and under 9 seconds on an RTX 4070.

API pricing undercuts the flagship

The BFL API serves the distilled checkpoint under the name flux-3-image-klein, priced per image by resolution: $0.008 at 768px, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K and $0.036 at 4K. The full FLUX 3 Image model charges $0.048 at 1K and $0.607 at 4K, which makes Klein roughly 5x cheaper at 1K and nearly 17x cheaper at 4K. Running the weights yourself drops the per-image cost to electricity.

A community license, not MIT

The weights are released under the FLUX Kommunity License: free commercial use for individuals and companies below $5M in annual revenue, along with research and non-commercial use. Larger organizations need a paid self-hosted license or the API. The guide is blunt that this is not MIT-style open source, though it is effectively free for freelancers, small studios and independent makers.

Quality: close to the leaders, with caveats

Black Forest Labs claims internal benchmarks beating Qwen-Image 2.1 and Ideogram V4.5 — the guide flags these as vendor claims that should be treated with caution. The one public third-party blind test it cites, covering 200 tasks and judged by Claude, Gemini and GPT, scores Nano Banana 2.1 at 7.20 versus 6.93 for Klein, with Klein at roughly one-fifth the cost per image.

Early testers report real weaknesses: generating people and poses is noticeably weak, multi-reference compositing gets shaky in complex scenes, and fine text rendering is mixed. Teams whose work centers on portraits should evaluate before switching.

Tooling available on day one

API users keep the same request format as FLUX 3 Image and only switch the endpoint. For local runs, Black Forest Labs shipped ComfyUI weights and a workflow at launch, plus Diffusers support, a Gradio bounding-box demo on GitHub and a try-it demo on Hugging Face Spaces. LoRA training works through Diffusers or the Ostris AI Toolkit on Klein Base.

Why it matters

Klein narrows the gap between hosted flagship models and what runs locally. If a 5.6B model on an 8GB card lands within a fraction of a point of leading closed models in blind testing, the practical trade-off changes: teams can generate at scale for the price of electricity, or through an API at a small fraction of flagship rates. The license preserves that freedom for smaller companies while routing larger ones toward paid terms, and the documented weaknesses around people, poses and fine text mean Klein is a strong general-purpose tool rather than a universal replacement.