FLUX.2 pushes image fidelity noticeably past SDXL, and it ships in two very different sizes. FLUX.2-dev is roughly 32B parameters: about 35 GB of fp8 weights plus a 12 GB text encoder, more than the total memory of a 12 GB consumer card, so it spills to SSD and generation slows to a crawl. Klein 4B is the distilled line in the same family — 4.07 GB of fp8 weights plus a 3.85 GB fp4 Qwen3-4B encoder, about 8.3 GB with the VAE.
What it needs: 8 GB for Klein, 24 GB+ for dev. Klein renders 1024x1024 in 4 steps: measured at 29 s on an 8 GB RTX 2060 Super, and 37 s for the 7.75 GB bf16 build on a 12 GB RTX 4070 Ti, the two outputs visually identical. Dev on that same 12 GB card took 32 min for one image at 24 steps, 15 min with an 8-step turbo LoRA, and 18 min with a Q3_K_S GGUF plus turbo — no quantization we tested makes dev usable at that size.
What it's good for: Klein for fast local stills on modest hardware; dev for high-fidelity work and character LoRA inference once a LoRA has been trained elsewhere. A LoRA binds to the variant it was trained on, so a dev-trained LoRA will not load into Klein.
In FlixML: `flux2_klein` (text-to-image on 8 GB) and `flux2_dev_lora` (dev text-to-image + trained character LoRA).