schema = "m64/capsule@1"
description = "MiniMax-H3-Image text-to-image on GB10 (diffusers). MiniMax's 33B open-weight image model (MiniMax-H3 family, guidance-distilled, 6 steps) with its Qwen3-VL-32B text encoder cut to the 51 layers it uses and quantized to 4-bit, at 2048x2048. MiniMax-H3-Image Community License: not licensed in the EU, UK, US or South Korea. Runs fully offline."
kinds = ["image.generate"]
authors = ["dh7"]
license = "other"

[hardware]
platform = "linux-aarch64"
gpu = "NVIDIA GB10"
gpu_count = 1
vram_gb = 88
driver = ">=580"
disk_gb = 140

[runtime]
sandbox = "container"
# No image: the Dockerfile builds on nvcr.io/nvidia/pytorch:25.09-py3 + diffusers 0.41.0, pinned by digest.
gpu = true
shm_gb = 8
internal_port = 8000
health = "/health"
expected_load_seconds = 2400
model_name = "minimax-h3-image"
user = "65534:65534"

[[artifacts]]
type = "hf"
repo = "MiniMaxAI/MiniMax-H3-Image"
revision = "5f90473023975e6271b0ed1dd84c4e88e4bbc6d3"
include = [
  ".gitattributes",
  "LICENSE",
  "README.md",
  "model_index.json",
  "modular_model_index.json",
  "processor/*",
  "scheduler/*",
  "text_encoder/*",
  "tokenizer/*",
  "transformer_diffusers/*",
  "vae_diffusers/*",
]
mount = "/weights/minimax-h3-image"

# The server's own defaults (minimax_h3_image_server.py), which every request gets: callers can only send
# prompt and seed. Returned as params_used.
[params]
size = "2048x2048"
num_inference_steps = 6

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