schema = "m64/capsule@1" description = "Z-Image 6B (base) text-to-image on GB10 (diffusers). Tongyi-MAI's undistilled 6B model (Qwen3-4B text encoder, Apache-2.0) at 1024x1024, 50 steps, CFG 4, bf16: slower than Turbo, more varied. Runs fully offline." kinds = ["image.generate"] authors = ["dh7"] license = "apache-2.0" [hardware] platform = "linux-aarch64" gpu = "NVIDIA GB10" gpu_count = 1 vram_gb = 32 driver = ">=580" disk_gb = 22 [runtime] sandbox = "container" # No image: the Dockerfile builds on vllm/vllm-openai (gemma4-cu130) + diffusers 0.37.1, pinned by digest. gpu = true shm_gb = 8 internal_port = 9012 health = "/health" expected_load_seconds = 300 model_name = "z-image-6b" user = "65534:65534" [[artifacts]] type = "hf" repo = "Tongyi-MAI/Z-Image" revision = "04cc4abb7c5069926f75c9bfde9ef43d49423021" include = [ ".gitattributes", "README.md", "model_index.json", "scheduler/*", "text_encoder/*", "tokenizer/*", "transformer/*", "vae/*", ] mount = "/weights/zimage" # The server's own defaults (z_image_server.py), which every request gets: callers can only send # prompt and seed. Returned as params_used. [params] size = "1024x1024" num_inference_steps = 50 guidance_scale = 4.0 [routes] extra = ["/v1/config"]