import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("CaaretAI/FLUX.2-klein-4B", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]FLUX.2-klein-4B (Caaret backup)
Backup of the public weights at black-forest-labs/FLUX.2-klein-4B.
These files are the original checkpoint. Caaret does not ship a requantized copy. Mining is a runtime hook: wide transformer linears (inner size ≥ 2048) run as noisy INT7 GEMM while the picture is drawn, and the same linears keep searching at 1024 rows when no picture is waiting. A real request takes the GPU before the next layer.
Official sampler for this distilled model: 1024×1024, 4 steps, guidance_scale=1.0.
On 2026-10-26 an idle pass of these weights on an RTX 3080 found Caaret block 180 (09c4e838e1102d0cce8f5fb055e85d0b253439a3f5decae7c4aa384a6ba7edd4).
License: Apache 2.0, same as the upstream model.
- Downloads last month
- 33
Model tree for CaaretAI/FLUX.2-klein-4B
Base model
black-forest-labs/FLUX.2-klein-4B