FLUX 3 Image Klein: a 5.6B open image model that runs on an 8GB GPU
Black Forest Labs put frontier-grade generation and pixel-perfect editing on local hardware — free under $5M revenue. Plus Nvidia's open 120B VL agent.
Copy markdown
Open weights, frontier quality, one consumer GPU
FLUX 3 Image Klein is out under open weights — 5.6B params, 11.2GB in BF16, and an INT8 build that fits in 8GB of VRAM. It does text-to-image, editing and multi-reference generation (up to 10 reference images) with native 4K output.
Pixel-perfect editing and bounding boxes
Bounding boxes place elements or confine edits to a defined region, and pixel restoration leaves untouched areas byte-identical — the real unlock for programmatic image pipelines. Day-one support across ComfyUI, Diffusers, vLLM, SGLang and TensorRT-LLM, with LoRA training via Ostris' AI Toolkit and NVIDIA NeMo.
~4s on a 5090, or $0.009 an image hosted
The distilled INT8 build renders a 1MP image in ~4s on an RTX 5090 and under 9s on a 4070; hosted, the BFL API runs $0.008–$0.036 per image, undercutting Nano Banana and GPT Image. Commercial use is free if your annual revenue is under $5M; above that you need a self-hosting license or the API.
Nvidia's Nemotron 3.5 Super VL goes open
Nvidia open-sourced a 120B-total / 12B-active multimodal MoE (hybrid Mamba-Transformer) under OpenMDW-1.1, with a 1M-token context and tool-calling. It reasons over screenshots, documents, charts and video — an open backbone for agents that need eyes, on Hugging Face, OpenRouter, ModelScope and NIM.