FLUX.2 [klein] 4B: about 3 seconds on a 12 GB card

Worked2026-10-24 · FLUX.2 [klein] 4B · RTX 4070 12GB, RTX 5070 12GB

FLUX.2 [klein] 4B (Apache-2.0) runs in 4 steps at cfg 1.

Results

• RTX 5070 12GB, 2 nodes: 3.3 s per 1024×1024 image (server-side median); 4.6–4.8 s from the client.
• RTX 4070 12GB, 2 nodes: 3.1 s (server-side); 4.7–5.0 s from the client.
• The 4070 and 5070 were within measurement noise of each other.

Notes

• Two nodes per GPU, so these are still reference values.
• The model, text encoder and VAE together are about 16 GB to download, which takes several minutes on a fresh node.