Llama-3.2-3B-MoE-4Expert

Llama-3.2-3B-MoE-4Expert

A 4-expert Mixture of Experts (MoE) model built on unsloth/Llama-3.2-3B-Instruct, specializing in chat, code, creative writing, and mathematics.

Model Description

This model combines four specialized Llama-3.2-3B variants into a single MoE architecture that intelligently routes queries to the most appropriate expert:

• General Chat & Explanations - Clear, structured responses for everyday queries
• Code & Programming - Python development, algorithms, and optimization
• Creative Writing - Stories, poetry, roleplay, and narrative content
• Mathematics - Equations, proofs, calculus, and step-by-step solutions

Architecture

• Base Model: unsloth/Llama-3.2-3B-Instruct
• Gate Mode: cheap_embed (embedding-based routing)
• Precision: bfloat16
• Merge Method: mergekit-moe
• Total Experts: 4

Expert Models

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Fu01978/Llama-3.2-3B-MoE-4Expert"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Write a Python function to calculate fibonacci numbers"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Performance

The model has been tested across all four domains and shows appropriate expert routing:

• General knowledge queries route to chat expert
• Programming tasks route to code expert
• Creative prompts route to creative expert
• Mathematical problems route to math expert

Edge cases and ambiguous queries default to the general chat expert.

Limitations

• 3B parameter size means less capable than larger models
• Expert routing depends on prompt phrasing
• May occasionally route to suboptimal expert for ambiguous queries
• Inherits limitations from constituent models

Acknowledgments

Built with mergekit. Thanks to:

• Meta AI for Llama 3.2
• unsloth for the base instruct model
• prithivMLmods for code and math variants
• DavidAU for the creative writing variant