Qwen3 MoE (Mixture of Experts) — 39M Parameters
Custom Qwen3 MoE model trained with pipeline parallelism.
Model Details
| Property | Value |
|---|---|
| Total Parameters | 39,388,928 |
| Architecture | MoE (Mixture of Experts) |
| Hidden Size | 128 |
| Num Layers | 2 |
| Attention Heads | 4 |
| Context Length | 512 |
| Vocab Size | 151,936 |
| Num Experts | 4 |
| Top-K Experts | 2 |
| MoE Hidden Dim | 128 |
Evaluation Results
| Metric | Value |
|---|---|
| val_loss | 0.3217 |
| val_perplexity | 1.3795 |
| train_loss | 0.2418 |
| step | 2000 |
Usage
import torch
from safetensors.torch import load_file
# Load model weights
state_dict = load_file("model.safetensors")
Training
Trained using pipeline parallelism with the multi_gpu_pretraining framework.
- Downloads last month
- 52