Qwen3-30M with GPT-2 Tokenizer (FP16)

A 30M parameter version of Qwen3-0.6B using GPT-2's tokenizer for better compatibility, optimized with FP16 precision.

Model Details

• Base Model: Qwen/Qwen3-0.6B
• Architecture: Qwen3 (8 layers, 224 hidden size)
• Parameters: ~35M (reduced from 637M)
• Tokenizer: GPT-2 (50,257 vocabulary)
• Vocabulary: Reduced from 151,936 to 50,257 tokens
• Precision: FP16 (half precision for memory efficiency)
• Model Size: ~60MB (vs ~120MB in FP32)

Architecture Specifications

• Layers: 8 transformer layers
• Hidden Size: 224
• Intermediate Size: 896 (4x hidden_size)
• Attention Heads: 8
• Key-Value Heads: 8
• Max Position Embeddings: 32,768
• Activation: SiLU

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load with automatic fp16 support
tokenizer = AutoTokenizer.from_pretrained("Mostafa8Mehrabi/qwen3-30m-fp16")
model = AutoModelForCausalLM.from_pretrained(
    "Mostafa8Mehrabi/qwen3-30m-fp16",
    torch_dtype=torch.float16,  # Explicitly use fp16
    device_map="auto"  # Automatically place on available device
)

# For GPU inference (recommended)
# model = model.to("cuda") # if you have a GPU

inputs = tokenizer("Hello, how are you?", return_tensors="pt")
# Move inputs to same device as model if using GPU
# inputs = {k: v.to(model.device) for k, v in inputs.items()}

outputs = model.generate(**inputs, max_length=50, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Key Features

• ✅ FP16 Precision: 50% smaller model size, faster inference
• ✅ 30M Parameters: Ultra-lightweight for edge deployment
• ✅ 8 Layers: Balanced depth for performance
• ✅ Standard GPT-2 tokenizer (no trust_remote_code)
• ✅ Compatible vocabulary sizes
• ✅ SafeTensors format for faster loading
• ✅ Works like any HuggingFace model
• ✅ 21x smaller than original Qwen3-0.6B
• ✅ GPU optimized for efficient inference

Architecture Comparison

Memory Requirements

• FP16: ~60MB model + ~30MB working memory = ~90MB total
• FP32: ~120MB model + ~60MB working memory = ~180MB total
• Memory savings: ~50% reduction compared to FP32
• Ultra-lightweight: Perfect for mobile and edge devices

Performance Notes

• FP16 provides significant memory savings with minimal quality loss
• 30M parameters optimized for fast inference while maintaining coherence
• Ideal for deployment in resource-constrained environments
• Compatible with both CPU and GPU inference
• Faster loading times due to smaller file size
• 8 layers provide good balance between model capacity and speed