zen-identity

Identity training for Zen AI models.

Zen Coder Family

Model	Base	Size	Script
zen-coder-4b	Zen Coder 4B	4B	train_4b.py
zen-coder	Zen Coder 24B base	24B	-
zen-coder-flash ⭐	Zen Coder Flash	31B MoE (3B active)	train_zen_coder_flash.py
zen-coder-max	Zen Coder MAX	671B MoE (14B active)	train_zen_coder_max.py

Training

zen-coder-flash (Flagship)

# CUDA
pip install torch transformers peft bitsandbytes datasets
python trainer/train_zen_coder_flash.py

# MLX (Apple Silicon)
git clone https://github.com/zenlm/zen-coder-flash
python training/train_mlx.py

zen-coder-max (Frontier)

Requires 4x A100 80GB or 8x H200:

python trainer/train_zen_coder_max.py

# Multi-GPU
torchrun --nproc_per_node 4 trainer/train_zen_coder_max.py

Contents

├── soul.md                      # Core identity document
├── datasets/                    # Training JSONL
├── trainer/
│   ├── train_zen_coder_flash.py # Zen Coder Flash training
│   ├── train_zen_coder_max.py   # Zen Coder MAX training
│   └── train_4b.py              # Zen Coder 4B training
└── training/
    └── app.py                   # HF Spaces Gradio app

Links

• Models: https://huggingface.co/zenlm
• Dataset: https://huggingface.co/datasets/zenlm/zen-identity
• zen-coder-flash repo: https://github.com/zenlm/zen-coder-flash