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_id | AMATERASU-32B-v0.1 |
| format | amaterasu-ckpt-v1 |
| frozen_total | 31,740,290,560 |
| freeze_hash | c1ff97b33d3f9280ccd5e066306a3ade4a23d8d7bbec484a3faa523a476129c2 |
| init dtype | fp32 shards |
| train dtype | bf16 |
Architecture
| d_model | 4096 |
| attention | GQA 32 / 8 / 128 |
| FFN | SwiGLU d_ff=11008 |
| HPT | 40 layers: Fast 12, Slow dense 8, Physical MoE 20 |
| MoE | 8 routed + 1 shared, top-2 |
| vocab | 65536, untied wte + lm_head |
| vision | 36-layer encoder, train 224×224, T_CLIP=16, N_CAM_MAX=6 |
| NCES | 6 layers, 128-d in, N_NODES_MAX=64 |
| clocks | Slow ~2–5 Hz, Fast ~30–100 Hz |
Parameter ledger
| component | params |
|---|---|
| vision_encoder | 1,605,837,824 |
| embeddings | 537,001,984 |
| shared_hpt_attn_norm | 1,677,895,680 |
| vision_ffns | 5,410,816,000 |
| language_ffns | 3,787,571,200 |
| physical_dense_ffns | 2,705,408,000 |
| physical_experts | 9,060,433,920 |
| agency_ffns | 3,787,571,200 |
| nces_encoder | 274,490,880 |
| audio_encoder | 59,879,424 |
| tactile | 17,825,792 |
| ssm | 428,032,000 |
| memory | 194,523,648 |
| latent_dynamics | 830,523,392 |
| eac_gcis | 799,404,544 |
| flow | 546,491,392 |
| ecd | 16,583,680 |
| TOTAL | 31,740,290,560 |
Files in this repo
| file | what |
|---|---|
README.md |
this card |
amaterasu_32b_v0.1.json |
freeze config (native, not Transformers config.json) |
manifest.json |
amaterasu-ckpt-v1 identity + weight map |
model.safetensors.index.json |
Hub shard index (same weight map). Not Transformers. |
*.safetensors |
51 freeze shards, ~119 GB fp32. nces.safetensors is Circuit-0 overlay |
metrics.json |
circuit-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:
ncesonly (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 fromamaterasu-ckpt-v1. - Overlay: trained NCES replaces freeze
nces.safetensors. Total stays31,740,290,560.
Load
Use AMATERASU resume_modules on this repo's 51 shards. Do not AutoModel.from_pretrained.