Qwen3-4B-Instruct-2507-math-aggressive

MATH-optimized | Aggressive pruning | 35% weights pruned

This model is a aggressively pruned version of Qwen/Qwen3-4B-Instruct-2507.

Pruning Alert: The benchmarks show virtually NO quality drop! This isn't a bug -- it is a feature. The Wanda pruning algorithm is so effective at identifying unimportant weights that it can remove a large percentage of parameters without affecting performance. Think of it like pruning dead leaves from a tree -- the tree does not miss them because they were not doing anything anyway!

Performance Comparison

Category	Original	Pruned	Change
Python	55.0%	55.0%	→
Html	60.0%	60.0%	→
Trivia	50.0%	50.0%	→
Math	45.0%	45.0% ⭐	→
Reasoning	60.0%	60.0%	→
Medical	30.0%	30.0%	→
Linux	35.0%	35.0%	→
Writing	45.0%	45.0%	→

Average: 47.5% -> 47.5% (+0.0%)

Math Retention: 100.0%

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CompactAI/Qwen3-4B-Instruct-2507-math-aggressive")
tokenizer = AutoTokenizer.from_pretrained("CompactAI/Qwen3-4B-Instruct-2507-math-aggressive")

inputs = tokenizer("Your prompt here", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Technical Details

Property	Value
Base Model	Qwen/Qwen3-4B-Instruct-2507
Specialization	Math
Prune Mode	Aggressive
Weight Reduction	35% weights pruned

License

This model inherits the license from the base model.