Black Forest Labs releases FLUX 3 Image Klein: 5.6B open-weights image generation and editing model with native 4K, bounding boxes, 8GB-VRAM INT8 build

Background

Open image models have forced a trade-off: quality came with 12B+ generators that strain local VRAM and are expensive to fine-tune, while small models lag on text rendering, local edits and multi-reference consistency. Teams that need on-prem or on-device editing often had to pick between quality and deployability.

Architecture

FLUX 3 Image Klein is the second open-weights FLUX 3 release after FLUX 3 Action, with a 5.6B generator. One model covers text-to-image, editing and multi-reference generation with up to 10 references, and keeps FLUX 3's native 4K, pixel-perfect editing and bounding boxes that pin element placement or confine edits. BFL says it draws on shared learnings across image, video and audio for physically grounded output. The release has a fine-tunable base, a step-distilled low-latency model, and FP8/INT8/NVFP4 builds made with NVIDIA.

Benchmarks (vendor-reported)

In BFL's internal evals it gains +91 Elo in text-to-image and +70 in editing over FLUX.2 [klein] 9B with 38% fewer parameters, and BFL ranks it ahead of Qwen-Image 2.1 (7B) and Krea 2 (12.9B). NVIDIA-measured speed: distilled + INT8 makes a 1MP image in ~4s on an RTX 5090 and under 9s on an RTX 4070; NVFP4 on GB300 runs under 0.6s. No independent leaderboard results yet.

Getting started

Pull weights from Hugging Face and run in ComfyUI or Diffusers; use the INT8 build on 8GB cards. The reference repo includes a Gradio demo for bounding-box control, and an HF Space is live. Fine-tune the base checkpoint with LoRA via Ostris AI Toolkit or NeMo AutoModel. In production, call flux-3-image-klein on the BFL API. Commercial use is free under $5M annual revenue per the FLUX Kommunity License; larger companies need self-hosted licensing or the API.