schema = "m64/capsule@1" description = "SenseNova-U1.5-8B-MoT text-to-image on GB10. SenseTime's native unified multimodal model (Qwen3-8B backbone with Mixture-of-Transformers generation, pixel-space, no VAE; Apache-2.0) generating 2048x2048 images with the official 8-step distillation LoRA merged: 8 steps, CFG 1. Runs fully offline." kinds = ["image.generate"] authors = ["dh7"] license = "apache-2.0" [hardware] platform = "linux-aarch64" gpu = "NVIDIA GB10" gpu_count = 1 vram_gb = 56 driver = ">=580" disk_gb = 40 [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 = 9204 health = "/health" expected_load_seconds = 1200 model_name = "sensenova-u1.5-8b-mot" user = "65534:65534" [[artifacts]] type = "hf" repo = "sensenova/SenseNova-U1.5-8B-MoT" revision = "9feeeab8a2792514d109cd34589342a2cc1d4ab2" include = [ ".gitattributes", "*.md", "*.json", "*.txt", "*.jinja", "*.safetensors", ] mount = "/weights/model" [[artifacts]] type = "hf" repo = "sensenova/SenseNova-U1.5-8B-MoT-LoRAs" revision = "f33b8fe0216e6f10e5aed58ebbfdd05f02826732" include = [".gitattributes", "README.md", "SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors"] mount = "/weights/loras" # The server's own defaults (sensenova_u1_server.py), which every request gets: callers can only send # prompt and seed. Returned as params_used. [params] size = "1:1" num_inference_steps = 8 guidance_scale = 1.0 timestep_shift = 3.0 cfg_norm = "none" think_mode = false [routes] extra = ["/v1/config"]