Black Forest Labs Open-Sources FLUX.3 Image Klein with Day-0 Ecosystem Support

Black Forest Labs officially released and open-sourced the image generation model FLUX.3 Image Klein (flux-3-image-klein-base) in mid-October. It quickly climbed Hugging Face's trending list, with inference frameworks and community tooling following up within the same day.

Confirmed

The model is an open-weights image model with 5.6B parameters; a single model handles text-to-image and multi-reference image-to-image generation, adds bounding box layout control, supports native 4K and pixel-level precise editing, and runs on a single consumer GPU. According to a post by @ariG23498, it generates an image in about 2 seconds on an H100.
The model centers on an image-to-image pipeline for generation and editing, is compatible with Hugging Face Diffusers (supported from day one of release), and its weights are openly available on Hugging Face.
The vLLM project announced Day-0 support for FLUX.3 Image Klein in vLLM-Omni, in partnership with BFL and NVIDIA, with a highlight being support for three compute precisions: BF16, FP8, and NVFP4.
The community ecosystem moved fast: user MrFakeName launched the Hugging Face Space "Klein Trainer," enabling no-code LoRA fine-tuning of FLUX.3 Klein on HF Jobs; per a repost, the community project Miles-Di… (original text truncated) has an associated open-source RL training pipeline that has been validated.

Why it matters

This is a significant iteration of BFL's open-source image models, integrating layout control and pixel-level editing into a single model, lowering the barrier to high-quality image generation and editing.
Day-one adoption by Diffusers and vLLM-Omni, along with community tools like no-code LoRA fine-tuning and an RL training pipeline, shows the model launched with a fairly complete open-source ecosystem, favoring rapid adoption and secondary development.

2026-10-15 ~ 2026-10-17 · 6 related posts

Full story(5 episodes)→

Episode 1: Black Forest Labs Open-Sources FLUX.3 Image Klein: 5.6B Model Runs 4K Generation on a Single Consumer GPU(2026-10-14, 35 posts)
Episode 2: Black Forest Labs Open-Sources FLUX.3 Image Klein with Day-0 Ecosystem Support(2026-10-15, 6 posts)
Episode 3: Community tests find FLUX 3 Klein and rivals regress on quality despite strong prompt adherence(2026-10-16, 4 posts)
Episode 4: Flux 3 Klein Tests Reveal Template Pitfalls and Mixed Editing Results(2026-10-16, 5 posts)
Episode 5: FLUX.3 Klein Tested: Mediocre Generation, Stellar High-Res Editing(2026-10-16, 2 posts)

Primary sources

Black Forest Labs open-sources FLUX.3 Image Klein: 2s per image on H100, layout control — ariG23498 · 2026-10-17 16:43 (UTC+08:00)
vLLM-Omni ships Day-0 serving for FLUX 3 Image Klein with BF16, FP8, NVFP4 support — vllm_project · 2026-10-15 23:27 (UTC+08:00)
unsloth ships FP8 quantized FLUX.3 klein for image editing on Hugging Face — unsloth · 2026-10-16 11:45 (UTC+08:00)

[source] vLLM-Omni ships Day-0 serving for FLUX 3 Image Klein with BF16, FP8, NVFP4 support — vllm_project · 2026-10-15
FLUX.3 open weights released with no-code LoRA finetuning on HF Jobs — realmrfakename · 2026-10-16
[source] unsloth ships FP8 quantized FLUX.3 klein for image editing on Hugging Face — unsloth · 2026-10-16
FLUX.3 Klein: 5.6B Open-Weights Image Model, RL Training Runs on One Node — ying11231 · 2026-10-17
[source] Black Forest Labs open-sources FLUX.3 Image Klein: 2s per image on H100, layout control — ariG23498 · 2026-10-17
Black Forest Labs' FLUX.3 Klein image model trends on Hugging Face — black-forest-labs · 2026-10-17