I finally got around to testing FLUX 3 Image Klein on my home setup after seeing the benchmarks. Running a 5.6B parameter model on an RTX 4070 felt like one of those things that shouldn't quite work yet.
It does work. The INT8 build fits in 8GB VRAM and generates a 1MP image in under 9 seconds through ComfyUI. That's not just a number — it means you can actually iterate on image generation locally during development without watching a progress bar and burning API credits.
What caught my attention wasn't just the speed. The model is open weights under the FLUX Kommunity License, which means free commercial use if you're under $5M revenue. Combined with LoRA fine-tuning support from day one, this changes what's realistic for small teams or indie projects. You can adapt the model on your own infrastructure, own your outputs, and avoid dependency on an API that might change pricing.
I'm not thinking "replace Midjourney" here. I'm thinking about what becomes viable when the inference runs on hardware you already own. Content tools, design automation, rapid prototyping workflows — things where round-trip latency kills the workflow when you're sending requests to a remote API.
Anyone else experimenting with local image generation for product ideas? I'm curious what use cases people are actually building around these smaller, capable models.