---
tags:
- text-generation
- reasoning
- coding
- mathematics
- quantization
license: apache-2.0
datasets:
- synthetic
base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
language:
- en
- hi
library_name: transformers
pipeline_tag: text-generation
---
# Alpie-Core: 4-bit Quantized Reasoning Model

📄 **[Technical Report: Alpie_Core.pdf](./Alpie_Core.pdf)**

## 1. Introduction

Alpie-Core is one of the world's first fine-tuned 4-bit reasoning models, proving that aggressive quantization can surpass full-precision baselines in reasoning, mathematics, and coding. By combining cutting-edge quantization-aware training with synthetic STEM-rich datasets, Alpie-Core achieves frontier-level reasoning while being practical for real-world deployment at scale.

## 2. Model Summary

- **Base Architecture**: DeepSeek-R1-Distill-Qwen-32B
- **Parameters**: 32 billion (quantized to 4-bit)
- **Training Method**: Supervised Fine-Tuning (SFT) using LoRA/QLoRA techniques
- **Quantization**: 4-bit NF4 with double quantization
- **Context Length**: 65k tokens
- **Max Output Length**: 16,384 tokens
- **License**: Apache 2.0


## 3. Approach

**Alpie-Core** has undergone extensive **supervised fine-tuning (SFT)** to strengthen reasoning, robustness, and safety. The training leveraged a diverse mixture of curated open-source datasets and proprietary synthetic data, optimized with high-quality LLM-generated responses. The fine-tuning process emphasized adherence to rigorous safety and usability standards, including:

1)**User Understanding and Clarity** – ensuring outputs are direct, interpretable, and pedagogically sound.

2)**Security and Ethical Guidelines** – filtering unsafe or harmful generations during and after training.

3)**Limitations, Disclaimers, and Knowledge Boundaries** – transparently communicating uncertainty and scope.

4)**Handling Complex and Sensitive Topics** – balancing informativeness with responsible guardrails.

5)**Safety and Respectful Engagement** – maintaining politeness, inclusivity, and cultural sensitivity.

6)**Confidentiality and Responsible Use** – preventing leakage of private training data, proprietary prompts, or internal reasoning traces.

This SFT approach enables Alpie-Core to deliver reliable, aligned, and context-aware responses while maintaining safety across a broad range of use cases.

## 4. Model Features

1. **Supports Streaming** – Real-time token-level responses
2. **OpenAI-Compatible API** – Seamless integration with OpenAI client libraries
3. **65K Context Length** – Handles very large inputs and conversations
4. **16,384 Max Output Length** – Enables extremely long generations
5. **4-Bit Quantization** – Memory-efficient and optimized for deployment
6. **High Throughput Inference** – Powered by vLLM for efficient large-scale serving
7. **Low Latency Inference** – Fast response times optimized for production
8. **Customizable Safety & Moderation Filters** – Built-in guardrails for safer outputs
9. **Supports Function Calling / Tool Use** – Enables structured outputs and external API integration

## 5. Key Highlights

1. **Frontier Performance in 4-bit**: 81.28% MMLU, 92.75% GSM8K, 57.8% SWE-Bench Verified

2) **STEM + Coding Excellence**: Outperforms full-precision peers in mathematics and programming

3) **Enhanced Content Access**: Provides factual responses to geopolitically sensitive topics

4) **Quantization Efficiency**: A 4-bit quantized variant achieves competitive performance retention compared to full-precision models, demonstrating that aggressive quantization can preserve task accuracy while substantially reducing hardware requirements.

5) **Benchmark Competitiveness**: Across more than ten standard evaluation benchmarks, the model demonstrates performance on par with or exceeding that of larger 70B+ parameter systems, highlighting the effectiveness of our training and optimization strategies.

6) **Environmental Benefits**: Through quantization and efficiency-focused design, the model requires significantly fewer computational resources. This translates into lower energy consumption and reduced carbon footprint relative to full-precision deployments.

## 6. Benchmark Results

![GSM8K Benchmark](Benchmark_GSM8K.png)

![AIME Benchmark](Benchmark_AIME.png)

![BBH Benchmark](BBH.png)

![Humanity's Last Exam](Humanity's_Last_Exam_(Text_Only)_-_Accuracy_Comparison.png)

![Combined Benchmark](combined_benchmark.png)

| Benchmark | Alpie-Core (32B-4bit) | DeepSeek-V2 (236B) | Qwen2.5 72B | Llama 3.1 405B | Llama 3.1 70B | Gemma-3 27B-PT | Mistral-Small-24B-Base-2501 |
|-----------|----------------------|-------------------|-------------|---------------|---------------|----------------|----------------------------|
| MMLU (5-shot) | **81.28%** | 78.4% | 85.0% | 84.4% | 79.3% | 78.6% | 80.73% |
| GSM8K (8-shot) | **92.75%** | 81.6% | 88.3% | 83.5% | - | 82.2% | 80.73% |
| BBH (3-shot) | **85.12%** | 78.8% | 79.8% | 82.9% | 81.6% | 77.7% | - |
| MMLU-Pro (5-shot) | **64.78%** | 51.4% | 58.3% | 52.8% | 53.8% | 52.2% | 54.37% |
| MBPP (pass@1) | **75.20%** | 65.0% | 72.6% | 68.4% | - | 65.6% | 69.64% |
| HumanEval (pass@1) | **57.23%** | 43.3% | 53.0% | 54.9% | - | 48.8% | = |

### SWE-Bench Verified Performance

![SWE-Bench Performance](swe.png)

| Rank | Model | Accuracy (%) | Performance vs Alpie |
|------|-------|-------------|---------------------|
| **1** | **Alpie Core** | **57.8** | **Alpie** |
| 2 | Qwen3-Coder-30B-A3B-Instruct | 51.6 | Below Alpie |
| 3 | o1 | 48.9 | Below Alpie |
| 4 | o3-mini (high) | 49.3 | Below Alpie |
| 5 | Claude 3.5 Sonnet | 49.0 | Below Alpie |
| 6 | DeepSeek R1 | 49.2 | Below Alpie |
| 7 | Devstral | 46.8 | Below Alpie |

### Humanity's Last Exam Leaderboard Performance

| Rank | Model | Accuracy (%) | Performance vs Alpie |
|------|-------|-------------|---------------------|
| 1 | GPT 4.5 Preview | 5.8 | Above Alpie |
| 2 | Claude Sonnet 4 | 5.42 | Above Alpie |
| **3** | **Alpie Core 32B (4-bit)** | **5.41** | **Alpie** |
| 4 | Llama 4 Maverik | 5.34 | Below Alpie |
| 5 | GPT 4.1 | 4.97 | Below Alpie |
| 6 | Kimi K2 Instruct | 4.68 | Below Alpie |
| 7 | DeepSeek V3 | 4.55 | Below Alpie |
| 8 | Gemini 1.5 Pro 002 | 4.55 | Below Alpie |

### Additional Benchmarks

| Benchmark | Alpie-Core (32B-4bit) | Category |
|-----------|----------------------|----------|
| AIME | **47.34%** | Advanced Mathematics |
| GPQA (Diamond) | **40.91%** | Graduate-level QA |
| TruthfulQA (MC2) | **60.05%** | Truthfulness |
| HellaSwag | **84.66%** | Commonsense |
| PIQA | **83.24%** | Physical Reasoning |
| ARC Challenge | **67.58%** | Science QA |
| CommonSenseQA | **87.06%** | Commonsense |
| AGIEval | **64.98%** | General Intelligence |
| Winogrande | **79.53%** | Commonsense Reasoning |
| MATH-500 | **70.00%** | Advanced Mathematics |

## 7. Training Details

- **Hardware**: 8× NVIDIA HOPPER-80GB GPUs
- **Fine-tuning Method**: LoRA/QLoRA with the following configuration:
  - LoRA Alpha: 16
  - LoRA Dropout: 0.05
  - LoRA Rank: 16
- **Quantization**: 4-bit NF4 + Double Quantization + FP16 compute
- **Dataset Domains**: Mathematics, coding, reasoning, science, general knowledge, competitive exams, Indian context + law, multilingual (Hindi and Hinglish)
- **Synthetic Data Advantage**: +15-20% performance boost in STEM & coding domains

## 8. Environmental Impact

![Carbon Footprint](carbon_footprint.png)

**Carbon Footprint**: We estimated the environmental impact of training Alpie-Core (32B) on 8× NVIDIA H100-80GB GPUs by calculating carbon emissions from GPU energy consumption. The calculation follows the formula:
CO₂e (kg) = Grid CO₂ Factor (kg/kWh) × Runtime (hours) × Power per GPU (kW) × Number of GPUs

Training Parameters:
Grid CO₂ Factor (Azure average): 0.364 kg CO₂e per kWh 
Runtime: 408 hours
GPUs: 8× H100-80GB
We report results under two assumption modes:

Realistic mode (average training draw ≈ 250 W per GPU = 0.25 kWh/hr): 0.364 × 408 × 0.25 × 8 ≈ 298 kg CO₂e


Conservative mode (near TDP ≈ 700 W per GPU = 0.70 kWh/hr): 0.364 × 408 × 0.70 × 8 ≈ 835 kg CO₂e


Total training footprint ranges from ~298 kg CO₂e (realistic) to ~835 kg CO₂e (conservative worst-case)




## 9. Use Cases

Best for **STEM**, **complex mathematical reasoning**, **coding**, and **Indian context**

1)**STEM**: Excels at solving advanced problems in science, technology, engineering, and mathematics with high accuracy.

2)**Complex Mathematical Reasoning**: Handles multi-step logical and quantitative reasoning tasks with strong reliability.

3)**Coding**: Supports software development, debugging, and algorithmic problem-solving across multiple programming languages.

4)**Indian Context**: Provides culturally aware insights, competitive exam assistance (JEE, NEET, UPSC), and multilingual support in Hindi/Hinglish.


## 10. Safety and Limitations

### Enhanced Content Access
Unlike the base DeepSeek model, Alpie-Core provides factual, balanced responses to geopolitically sensitive questions, offering global accessibility and factual accuracy on topics like Taiwan's status, Arunachal Pradesh sovereignty, and other sensitive geopolitical issues.

### Current Limitations
- Multilingual reasoning in Hindi/Hinglish shows room for improvement
- Fixed knowledge cutoff without real-time information retrieval
- Occasional struggles with complex multi-hop mathematical reasoning
- Potential hallucinations in factual question-answering

### Mitigations
- Safety classifiers and output filtering systems
- Model-assisted safety pipeline using RLHF
- Comprehensive adversarial testing by domain experts

## 11. How to Use

### Non-Streaming Inference
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, PeftConfig
import torch

# Load LoRA adapter configuration to find the base model
peft_model_id = "169Pi/Alpie-Core-4-bit"
config = PeftConfig.from_pretrained(peft_model_id)

# Load the base model
base_model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load LoRA weights
model = PeftModel.from_pretrained(base_model, peft_model_id)

# Ensure evaluation mode
model.eval()

# Sample inference
prompt = "Solve the Riemann Hypothesis and provide a final answer?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=1000)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print("Response:\n", response)
```

### Streaming Inference
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
from peft import PeftModel, PeftConfig
import torch

# Load LoRA adapter configuration to find the base model
peft_model_id = "169Pi/Alpie-Core-4-bit"
config = PeftConfig.from_pretrained(peft_model_id)

# Load the base model
base_model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load LoRA weights
model = PeftModel.from_pretrained(base_model, peft_model_id)

# Ensure evaluation mode
model.eval()

# Initialize streamer
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

# Sample streaming inference
prompt = "Solve the Riemann Hypothesis and provide a final answer?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

print("Streaming Response:")
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1000,
        streamer=streamer,
        do_sample=True,
        temperature=0.7,
        top_p=0.9
    )
```

### Deployment Options
- **Transformers**: Python, PyTorch integration
- **vLLM**: High-throughput inference

## 12. Citation

```bibtex
@misc{alpie2025core,
  title     = {Alpie-Core: A 4-bit Quantized Reasoning Model Surpassing Full-Precision Benchmarks},
  author    = {Alpie AI},
  year      = {2025},
  url       = {https://huggingface.co/alpie/Alpie-Core-4bit}
}
```

## 13. License

Apache 2.0 – Free for research and commercial use

## 14. Acknowledgements / Credits

We would like to thank **DeepSeek** for their original model, which served as the foundation for this work. Our team fine-tuned the model and implemented **4-bit quantization**, achieving improved efficiency and accuracy for downstream tasks. This model is built with respect to the contributions of the original authors and aims to provide a safe, high-performance solution for reasoning and inference.

## 15. Contact

For technical inquiries and support: **contact@169pi.com**

---

*For technical details, training methodology, and comprehensive evaluation results, please refer to our technical report.*