nanokimi-mini

This repository contains the nanoKimi model pre-trained on Shakespeare dataset. An upgraded version of nanokimi trained on OpenWebText will be up on HuggingFace in a few days.

Model Details

• Architecture: 12 layers, 12 heads, 768 embedding dimension
• Training Data: Shakespeare dataset
• Features: Mixture of Experts (8 experts), Latent Attention
• Model Type: Kimi-K2

Files

• pytorch_model.bin - Model weights
• config.json - Model configuration
• src/ - Source code for model architecture
• modeling_kimik2.py - HuggingFace wrapper

Usage

import torch
import json
from huggingface_hub import hf_hub_download

# Download files
config_path = hf_hub_download(repo_id="sohv/nanokimi-mini", filename="config.json")
weights_path = hf_hub_download(repo_id="sohv/nanokimi-mini", filename="pytorch_model.bin")

# Load config and weights
with open(config_path) as f:
    config = json.load(f)

weights = torch.load(weights_path, map_location="cpu")
print("Model downloaded successfully!")

License

MIT License

Contact

Raise an issue in Files and Version or reach out to me here for any feedback or enquiry.