MonumentalSystems/wire-2m-chinchilla

Wire-2M (H16 L8, Chinchilla)

WireNative 2M — Chinchilla-scaled training run, best BPB 1.97

Part of the Harmonic GPT research into oscillator-based neural computation.

Architecture: WireNative

Property	Value
Parameters	2,326,392
BPB	1.9697
Training step	0
n_harmonics	32
n_layers	8
n_groups	7
d_model	448
Vocab	256 (raw bytes)

Usage

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

weights = load_file(hf_hub_download("MonumentalSystems/wire-2m-chinchilla", "model.safetensors"))
config = json.load(open(hf_hub_download("MonumentalSystems/wire-2m-chinchilla", "config.json")))

All operations are native Clifford algebra / harmonic oscillator dynamics — no softmax attention, no MLP, no ReLU.