---
license: apache-2.0
datasets:
- unsloth/OpenMathReasoning-mini
- mlabonne/FineTome-100k
language:
- en
base_model:
- Qwen/Qwen3-4B
pipeline_tag: text-generation
library_name: transformers
tags:
- moe
- math
- code
- text-generation-inference
- trl
---
![Draconis.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/CpdC-5a9DZO7NMY6DcW9M.png)

# Draconis-Qwen3\_Math-4B-Preview

> **Draconis-Qwen3\_Math-4B-Preview** is fine-tuned on the **Qwen3-4B** architecture, optimized for excellence in **mathematical reasoning**, **logical problem solving**, and **structured content generation**. This preview model focuses on precision, step-by-step reasoning, and efficient inference, making it ideal for educational and technical applications where reliability and compact performance are essential.

> [!note]
GGUF [Q4_K_M] : https://huggingface.co/prithivMLmods/Draconis-Qwen3_Math-4B-Preview-Q4_K_M-GGUF

> [!note]
GGUF [Q5_K_M] : https://huggingface.co/prithivMLmods/Draconis-Qwen3_Math-4B-Preview-Q5_K_M-GGUF

## Key Features

1. **Mathematical and Logical Reasoning**
   Finetuned to solve symbolic logic, arithmetic, and multi-step mathematical problems, making it ideal for STEM learning, competitions, and educational use.

2. **Compact Code Understanding**
   Efficient in writing and interpreting code in Python, JavaScript, and other languages, suitable for lightweight coding tasks and algorithmic explanations.

3. **Factual Precision**
   Trained on high-quality, curated data with reasoning benchmarks to reduce hallucinations and ensure correctness in technical outputs.

4. **Instruction-Tuned**
   Strong adherence to instructions, ideal for structured queries, step-by-step problem solving, and producing formatted outputs (Markdown, JSON, tables).

5. **Multilingual Support**
   Capable of understanding and responding in over 20 languages, useful for multilingual education and technical translation.

6. **Efficient Performance**
   Based on the 4B parameter variant of Qwen3, optimized for resource-constrained environments without compromising core reasoning capability.

## Quickstart with Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Draconis-Qwen3_Math-4B-Preview"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Solve the equation: 3x + 7 = 22. Show all steps."

messages = [
    {"role": "system", "content": "You are a step-by-step math tutor."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
```

## Intended Use

* Solving math and logic problems
* Code assistance and basic debugging
* Education-focused applications (STEM tutoring)
* Structured content generation (e.g., JSON, Markdown)
* Multilingual reasoning and translations
* Lightweight deployment in reasoning tasks

## Limitations

* Limited creativity in open-ended or fictional content
* May struggle with ambiguous or multi-intent prompts
* Smaller context window compared to 14B+ variants
* Still subject to factual errors in edge cases or adversarial queries

## References

1. [AIMO-2 Winning Solution: Building State-of-the-Art Mathematical Reasoning Models] :  https://arxiv.org/pdf/2504.16891
2. [YaRN: Efficient Context Window Extension of Large Language Models] : https://arxiv.org/pdf/2309.00071