FLUX 3 Image Klein: open 4K image model runs on an 8GB GPU
Black Forest Labs puts frontier image gen on a consumer GPU, free for small shops — while NVIDIA ships 120B open Nemotron VL plus edge Bolt models.
Copy markdown
4K image generation on the GPU you already own
FLUX 3 Image Klein is a 5.6B open-weights model (38% smaller than FLUX.2 klein) that does text-to-image, pixel-perfect editing, and generation from up to 10 reference images — natively at 4K. An INT8 build runs on an 8GB-VRAM NVIDIA card and renders a 1MP image in under 9 seconds on an RTX 4070. Weights are on Hugging Face with day-one ComfyUI, Diffusers, and LoRA-training support, so a local frontier image pipeline is a weekend project now.
Free to ship — if you earn under $5M
BFL claims the top Elo among open image models, ahead of FLUX.2 pro and Ideogram V4.5, with bounding-box placement and pixel-accurate edits. The catch is the FLUX Kommunity License: commercial use is free only for companies under $5M/year revenue; above that you license it or use the hosted API at $0.009 per 1,000 images.
Nemotron 3.5 Super VL: 120B open, with the data and recipes
NVIDIA shipped a 120B-total / 12B-active MoE that reads text, images, documents, and video in one model — tuned for OCR, tables, charts, and PDFs. Unusually, it opens the weights, ~45 vision-language datasets, and the training recipes, so you can actually reproduce or re-tune it. It's live on Hugging Face, OpenRouter, and ModelScope, with an NVFP4 4-bit checkpoint for Blackwell.
Bolt: a 24GB agent model for the Spark laptop
Alongside Super VL, NVIDIA's Bolt 74B-A9B quantizes natively to 2-bit to fit in 24GB and runs on llama.cpp, Ollama, LM Studio, and vLLM — aimed at the RTX Spark laptops landing now. A tiny Bolt E4B for 4GB devices (Jetson, embedded voice agents) follows Oct 23.