MonumentalSystems/wire-9m-best

Wire-9M (H64 L8)

WireNative 9M — 64 harmonics, 8 layers, selected BPB checkpoint.

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

Architecture: WireNative

Property	Value
Parameters	8,894,520
BPB	3.0866
Training step	5,000
n_harmonics	64
n_layers	8
n_groups	7
d_model	896
Vocab	256 (raw bytes)

Usage

This repository contains weights and configuration for research inspection. It does not currently bundle a standalone transformers runtime.

import json

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

weights = load_file(hf_hub_download("MonumentalSystems/wire-9m-best", "model.safetensors"))
config = json.load(open(hf_hub_download("MonumentalSystems/wire-9m-best", "config.json")))

Sequence mechanism

This is an attention-free oscillator model. Its sequence mixing uses native Clifford algebra and harmonic dynamics rather than softmax attention, an attention KV cache, MLPs, or ReLU activations. The reported BPB is a run-selection diagnostic and should not be compared across corpora or evaluation protocols without reproducing the original setup.