dqnCode-v1/README.md

---
license: apache-2.0
language:
- en
tags:
- code
- coding
- programming
- reasoning
- small-model
- efficient
- local
- qwen
- qwen3
- qwen3.5
- 4b
- small
- developer
- coding-assistant
- python
- debugging
- daily-use
- localai
- ai
- gpt
- dqnlabs
- dqngpt
- gguf
- lmstudio
- ollama
pipeline_tag: text-generation
---

# dqnCode-v1

dqnCode-v1 is a 4B-parameter language model designed for fast, clear, and practical coding assistance.

It focuses on writing, fixing, and explaining code efficiently, with minimal verbosity and strong real-world usefulness. It is optimized for everyday programming tasks with low latency and concise outputs.

![dqnCode Banner](dqnCode.png)
---

## Benchmark

dqnCode-v1 is positioned as a high-performance compact coding model, with strong results on standard code generation benchmarks. It is trained with simple prompts in mind, so you don't need to be a developer to use it!

### HumanEval

- **pass@1:** 63.4%  

This score places dqnCode-v1 among the strongest models in the 4B parameter class for coding tasks (only beaten by one other model in the 4B or below models class!)

| Model                     | Provider        | HumanEval (pass@1) |
|--------------------------|-----------------|--------------------|
| GPT-3.5 Turbo            | OpenAI          | 68%                |
| GPT-4                    | OpenAI          | 67%                |
| dqnCode v1 (4B)          | DQN Labs        | 63.4%              |
| Phi-3.5-mini-instruct    | Microsoft       | 62.8%              |
| DeepSeek Coder 33B       | DeepSeek        | 52.4%              |
| Gemma 2 27B              | Google          | 51.8%              |
| Nous Hermes 3 405B       | Nous Research   | 51.4%              |
---

## Benchmark Context

- Evaluated on HumanEval (Python code generation benchmark)  
- Focused on functional correctness of generated code  
- Designed to reflect real-world coding performance in a compact model  

---

## Positioning

dqnCode-v1 is designed to deliver coding performance comparable to significantly larger models while maintaining a small footprint.

- Approaches GPT-3.5 / GPT-4-level performance on many practical coding tasks  
- Achieves strong performance with significantly fewer parameters  
- Optimized for fast local inference and real-world development workflows
- Excellent at all languages, especially HTML-CSS-JS, Python, Java, and more!


## Model Description

- **Model type:** Causal Language Model  
- **Parameters:** 4B  
- **Primary use:** Coding and programming assistance  
- **Style:** Direct code generation with minimal explanation  

dqnCode-v1 4B is optimized for practical coding workflows rather than long-form development or benchmark performance.

---

## Intended Uses

### Direct Use

- Writing code snippets  
- Debugging errors  
- Explaining code logic  
- Converting code between languages  
- Assisting with small to medium programming tasks  
- Competitive programming support  

---

## Key Characteristics

- Produces clean and functional code  
- Prioritizes correctness and simplicity  
- Minimizes unnecessary explanation  
- Strong at debugging and fixing issues  
- Designed for consistent outputs across common coding tasks  
- Excellent instruction following and stability.
- Trained with simple prompts in mind, so you don't need to be a developer to use it!

---

## Example

**Input**
```text
Write a Python function to check if a number is prime
```

**Output**
```python
def is_prime(n):
    if n < 2:
        return False
    for i in range(2, int(n**0.5) + 1):
        if n % i == 0:
            return False
    return True
```

---

**Input**
```text
Fix this code:
for i in range(5)
    print(i)
```

**Output**
```python
for i in range(5):
    print(i)
```

---

## Usage

This model is available on many platforms and is compatible with many formats.

The GGUF format is compatible with llama.cpp and LM Studio.  
Other formats include MLX (LM Studio, optimized for Apple devices), and HF (universal compatibility).

---

## Training Details

dqnCode-v1 is fine-tuned for practical coding tasks and efficient problem solving.

The training process emphasizes:

- Functional correctness  
- Minimal and clean outputs  
- Real-world coding scenarios  
- Debugging and code repair  

---

## Limitations

- Limited performance on very large or complex codebases  
- Not optimized for long-form software architecture design  
- May simplify explanations rather than provide deep theoretical detail  

---

## Efficiency

dqnCode-v1 is designed to run efficiently on consumer hardware, with support for quantized formats.

---

## License

Apache 2.0

---

## Author

Developed by DQN Labs.  
This model card was generated with the help of dqnGPT v0.2!