schema = "m64/capsule@1"
description = "FLUX.2 [klein] 4B text-to-image on GB10 (diffusers). Black Forest Labs' distilled 4B rectified-flow model (Qwen3-4B text encoder, Apache-2.0) at 1024x1024, 4 steps, guidance 1, bf16. A small, fast FLUX.2."
kinds = ["image.generate"]
authors = ["dh7"]
license = "apache-2.0"

[hardware]
platform = "linux-aarch64"
gpu = "NVIDIA GB10"
gpu_count = 1
# ~8 GB transformer + ~8 GB Qwen3-4B text encoder (bf16) + VAE and activations.
vram_gb = 24
driver = ">=580"
disk_gb = 18

[runtime]
sandbox = "container"
# No image: the Dockerfile builds on nvcr.io/nvidia/pytorch:25.09-py3 pinned by digest.
gpu = true
shm_gb = 8
internal_port = 9019
health = "/health"
expected_load_seconds = 300
model_name = "flux2-klein-4b"
user = "65534:65534"

[[artifacts]]
type = "hf"
repo = "black-forest-labs/FLUX.2-klein-4B"
revision = "e7b7dc27f91deacad38e78976d1f2b499d76a294"
# The diffusers folders only: not the 7.7 GB single-file (ComfyUI) checkpoint, not the
# card's images.
include = [
  ".gitattributes", "LICENSE.md", "README.md", "model_index.json",
  "scheduler/*", "text_encoder/*", "tokenizer/*", "transformer/*", "vae/*",
]
mount = "/weights/flux2"

[params]
size = "1024x1024"
num_inference_steps = 4
guidance_scale = 1.0

[routes]
extra = ["/v1/config"]