Verified benchmarks (H200 GPU, Qwen models):

Model File Size VRAM (Inference) VRAM (+ Training) Speed 7B 5.57 GB 5.67 GB ~8 GB 37.2 tok/s 14B 9.95 GB 10.08 GB ~14 GB 20.8 tok/s 32B 19.23 GB 19.47 GB ~26 GB 10.9 tok/s 72B 41.21 GB 41.54 GB ~52 GB 6.3 tok/s What this means:

Train 7B models on RTX 3070/4070 (8GB) Train 32B models on RTX 3090/4090 (24GB) Train 70B models on A100-40GB or 2x 3090

Usage: from zse.format import load_zse_model from zse.training import LoRAConfig, add_lora_to_model

model, tokenizer, info = load_zse_model("model.zse") model = add_lora_to_model(model, LoRAConfig(rank=16, alpha=32))

# Train normally, adapter is ~25MB save_lora_adapter(model, "my_adapter.safetensors") Trainable params: 0.2% of model (12M params for 7B)

pip install zllm-zse[training]

Code: github.com/zyora-ai/zse