FLUX.3 Klein: Black Forest Labs ships a 5.6B open image model
The distilled checkpoint does 4K text-to-image and multi-ref edits locally, and BFL claims wins over FLUX.2 pro; Fish Audio's open drama-3 TTS also dropped.
Copy markdown
A local image editor you actually own
FLUX.3 Klein is a 5.6B open-weights model doing text-to-image, single- and multi-reference editing (up to 10 reference images), bounding-box layout control, and native 4K output. That's a self-hostable stand-in for a hosted editor — build a product mockup pipeline or a batch asset generator with no per-image API bill.
It fits the GPU you already have
The released checkpoint is distilled to 16 steps with no guidance. Official NVFP4 weights peak around 25GB, but INT8 via ComfyUI reportedly runs on 8GB cards — about 4 seconds per image on an RTX 5090 and under 9 on a 4070. Local iteration speed, not a cloud queue.
The catches: a 9B upsampler, shaky people
BFL warns “do not use the model without a prompt upsampler” — a separate 9B model that adds ~28GB. Early users also report weak fine-text rendering and poor human anatomy (NSFW appears stripped from training), with MOSS-Image still ahead on text-in-image. Budget VRAM for the upsampler and test faces before you ship.
License: free to ship if you're under $5M
The FLUX Kommunity License v1.0 permits commercial output generation for users under $5M in annual revenue, plus non-commercial use for everyone. An undistilled “Klein Base” checkpoint also ships for fine-tuning and research, so you can specialize it on your own style.
Also open today: Fish Audio drama-3
A 35B mixture-of-experts TTS model (3B active) covering 83 languages, up to 16 characters per script, inline laugh/sigh tags, voice cloning, and audio editing. It hits ~89ms median first-audio on an H200 and runs faster than real time up to 64 concurrent streams — serve it with vLLM-Omni or SGLang-Omni for a self-hosted voice stack.