CoolFace
Modelpublic

MagistrTheOne/NULLXES-AMATERASU

sourceHugging Faceapache-2.0updated 23d agoView on Hugging Face
0likes
Model Card

NULLXES AMATERASU-32B v0.1

Author: MagistrTheOne | NULLXES Hub: MagistrTheOne/NULLXES-AMATERASU Code: github.com/MagistrTheOne/NULLXES-AMATERASU

From-scratch Embodied Agency Foundation Model. Native format amaterasu-ckpt-v1. Not Transformers, not π0 / GR00T / OpenVLA / Qwen / Llama weights. Do not AutoModel.from_pretrained.

Identity

model_idAMATERASU-32B-v0.1
formatamaterasu-ckpt-v1
frozen_total31,740,290,560
freeze_hashc1ff97b33d3f9280ccd5e066306a3ade4a23d8d7bbec484a3faa523a476129c2
init dtypefp32 shards
train dtypebf16

Architecture

d_model4096
attentionGQA 32 / 8 / 128
FFNSwiGLU d_ff=11008
HPT40 layers: Fast 12, Slow dense 8, Physical MoE 20
MoE8 routed + 1 shared, top-2
vocab65536, untied wte + lm_head
vision36-layer encoder, train 224×224, T_CLIP=16, N_CAM_MAX=6
NCES6 layers, 128-d in, N_NODES_MAX=64
clocksSlow ~2–5 Hz, Fast ~30–100 Hz

Parameter ledger

componentparams
vision_encoder1,605,837,824
embeddings537,001,984
sharedhptattn_norm1,677,895,680
vision_ffns5,410,816,000
language_ffns3,787,571,200
physicaldenseffns2,705,408,000
physical_experts9,060,433,920
agency_ffns3,787,571,200
nces_encoder274,490,880
audio_encoder59,879,424
tactile17,825,792
ssm428,032,000
memory194,523,648
latent_dynamics830,523,392
eac_gcis799,404,544
flow546,491,392
ecd16,583,680
TOTAL31,740,290,560

Files in this repo

filewhat
README.mdthis card
amaterasu_32b_v0.1.jsonfreeze config (native, not Transformers config.json)
manifest.jsonamaterasu-ckpt-v1 identity + weight map
model.safetensors.index.jsonHub shard index (same weight map). Not Transformers.
*.safetensors51 freeze shards, ~119 GB fp32. nces.safetensors is Circuit-0 overlay
metrics.jsoncircuit-0 run metrics

No Transformers config.json. Hub Parameters chip reads model.safetensors.index.json + shard headers. Source of truth is still frozen_total above.

Circuit-0 (this release)

  • —Data: HiFi-UMI-2K parquet only (CC BY 4.0). No video. No intent_label (hand motion ≠ ACT).
  • —Trainable: nces only (274,490,880). Rest of freeze graph frozen.
  • —Hardware: 1× NVIDIA H200 141 GB, bf16 autocast. Not full 32B AdamW.
  • —Gate: finite L_mm, checkpoint write, resume from amaterasu-ckpt-v1.
  • —Overlay: trained NCES replaces freeze nces.safetensors. Total stays 31,740,290,560.

Load

Use AMATERASU resume_modules on this repo's 51 shards. Do not AutoModel.from_pretrained.