Trained with a focus on strict structural discipline, zero-waffle response generation, and rapid inference execution, Sparky delivers exceptional performance across code generation, agentic tool workflows, and mathematical reasoning within a lightweight sub-5B parameter footprint.

Based on the renowned Qwen2.5-3B base architecture.

Designed for coding, instruction following and assistant tasks normal 4B models cant.

Benchmark Suite

Evaluated using lm-evaluation-harness over local API completion endpoints:

Model	Benchmark	Category	Score	Evaluation Setting
Qwen 2.5 3B	HumanEval	Python Code Generation	42.10%	0-Shot (`pass@1`)
Sparky 4B V1	HumanEval	Python Code Generation	51.22%	0-Shot (`pass@1`)
Qwen 2.5 3B Instruct	IFEval	Instruction Following	42.50%	0-Shot (Strict)
Sparky 4B V1	IFEval	Instruction Following	47.32%	0-Shot (Strict)
Qwen 2.5 3B	GSM8K	Multi-Step Math Reasoning	79.10%	5-Shot (Flexible Extract)
Sparky 4B V1	GSM8K	Multi-Step Math Reasoning	61.03%	5-Shot (Flexible Extract)

Qwen have not released non instruct benchmarks of IFEval offically source https://qwen.ai/blog?id=qwen2.5-llm

How to install

To run Sparky locally, we recommend using Ollama: ollama run hf.co/Noid3a-Labs/Sparky-4B-V1:Q4_K_M

You can also install a Q4 or Q3 varient manually from the Files and versions tab on the top navbar

Prompt Format (ChatML)

Sparky utilizes standard ChatML formatting:

text

<|im_start|>system
You are Sparky, a helpful and precise AI assistant built by Noid3a Labs.<|im_end|>
<|im_start|>user
Write a Python function to check if a string is a palindrome.<|im_end|>
<|im_start|>assistant

This model is a derivative of Qwen and is subject to the Qwen Research License Agreement.