README.md

---
license: mit
base_model:
- Qwen/Qwen3.5-4B
tags:
- code
---
---

## license: mit

<div align="center">

# 🧠 Nexora-Qwen-Coder-4B

### Compact Agentic Coding Model

**Coding · Debugging · Tool Use · Structured Reasoning · Local AI**

<br>

[![Hugging Face](https://img.shields.io/badge/Hugging%20Face-Nexora--Qwen--Coder--4B-yellow?logo=huggingface)](https://huggingface.co/guell00/Nexora-Qwen-Coder-4B)
[![GGUF](https://img.shields.io/badge/Format-GGUF-blue)](https://github.com/ggerganov/llama.cpp)
[![Unsloth](https://img.shields.io/badge/Fine--Tuning-Unsloth-orange)](https://unsloth.ai/)
[![License](https://img.shields.io/badge/License-MIT-lightgrey)](#-license)

</div>

<br>

<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/644afe169279988e0cbcd2d9/AVI5Rk73jQfx76ePPXx90.png" width="100%">
</p>

---

> **Nexora-Qwen-Coder-4B** is a compact, coding-focused language model fine-tuned from the **Qwen 3.5 4B family**, with an emphasis on **code generation, debugging, structured reasoning, tool use, and local agentic workflows**.

The core idea behind Nexora is simple:

> **A small coding model should do more than generate code. It should understand the task, reason through problems, interact with tools, inspect feedback, and iterate toward a solution.**

Designed for developers who want capable AI assistance without requiring a large datacenter-scale deployment.

---

# ✦ Model Overview

| Property             | Details                        |
| :------------------- | :----------------------------- |
| **Model**            | `Nexora-Qwen-Coder-4B`         |
| **Hugging Face**     | `guell00/Nexora-Qwen-Coder-4B` |
| **Base Model**       | Qwen 3.5 4B                    |
| **Architecture**     | Dense Transformer              |
| **Model Class**      | 4B Parameters                  |
| **Primary Focus**    | Coding & Agentic Workflows     |
| **Fine-Tuning**      | Nexora Fine-Tuning             |
| **Training Method**  | SFT + Curriculum Learning      |
| **Reasoning Data**   | Trace Inversion                |
| **Agent Data**       | Tool-Use & Agent Trajectories  |
| **Training Context** | Up to 32K tokens               |
| **Evaluation**       | MTP, `n=2`                     |
| **Format**           | GGUF                           |
| **Inference**        | llama.cpp compatible           |

---

# ⚡ Why Nexora?

Nexora-Qwen-Coder-4B is not designed around parameter count alone.

The objective is to make a **compact local model more useful for real software development workflows**.

The fine-tuning strategy focuses on four primary capabilities:

### `01` · Coding

Generate, complete, explain, refactor, and implement code across common programming tasks.

### `02` · Debugging

Analyze errors, identify potential failure points, reason about bugs, and produce targeted fixes.

### `03` · Agentic Workflows

Operate in environments where the model can inspect files, select tools, execute actions, receive feedback, and iterate.

### `04` · Structured Reasoning

Handle multi-step technical tasks that benefit from planning, decomposition, and structured problem solving.

---

# 🎛️ Recommended Sampling

For the best balance of **consistency, coding quality, and controlled generation**, the recommended default configuration is:

| Parameter              | Recommended |
| :--------------------- | :---------: |
| **Temperature**        |    `0.1`    |
| **Top P**              |    `0.95`   |
| **Top K**              |     `52`    |
| **Min P**              |    `0.05`   |
| **Repetition Penalty** |    `1.1`    |
| **Presence Penalty**   |    `Off`    |

### Default Configuration

```text
Temperature:        0.1
Top P:              0.95
Top K:              52
Min P:              0.05
Repetition Penalty: 1.1
Presence Penalty:   Off
```

This configuration is recommended for:

* Code generation
* Code completion
* Debugging
* Refactoring
* Technical reasoning
* Tool calling
* Agentic coding workflows

The low `Temperature` is intended to improve consistency while preserving a small amount of generation flexibility.

> **Note:** Evaluation results may vary when using sampling parameters different from those used during benchmarking.

---

# 📊 Benchmark Performance

Nexora-Qwen-Coder-4B was evaluated locally using the **benchlocal** evaluation framework.

The evaluation focuses primarily on practical developer workflows, including debugging, tool use, agent behavior, and instruction following.

## Results

| Benchmark             | Nexora-Qwen-Coder-4B | Qwen 3.5 4B |   Delta |
| :-------------------- | -------------------: | ----------: | ------: |
| **BugFind-15**        |         **71 / 100** |    52 / 100 | **+19** |
| **HermesAgent-20**    |         **64 / 100** |    61 / 100 |  **+3** |
| **ToolCall-15**       |        **100 / 100** |    90 / 100 | **+10** |
| **InstructFollow-15** |         **93 / 100** |    93 / 100 |   **0** |

### Relative Evaluation Snapshot

```text
BugFind-15
Nexora-Qwen-Coder-4B  ██████████████░░░░░░  71
Qwen 3.5 4B           ██████████░░░░░░░░░░  52

HermesAgent-20
Nexora-Qwen-Coder-4B  █████████████░░░░░░░  64
Qwen 3.5 4B           ████████████░░░░░░░░  61

ToolCall-15
Nexora-Qwen-Coder-4B  ████████████████████ 100
Qwen 3.5 4B           ██████████████████░░  90

InstructFollow-15
Nexora-Qwen-Coder-4B  ██████████████████░░  93
Qwen 3.5 4B           ██████████████████░░  93
```

### Key Observations

The strongest observed results were in:

* **Debugging**
* **Tool calling**
* **Coding-oriented workflows**
* **Instruction following**
* **Local agent scenarios**

These results suggest that the fine-tuning process improved the model's performance on targeted coding and agentic tasks compared with the base evaluation reference.

> Benchmark results are snapshots from specific evaluation runs. They should not be interpreted as universal performance guarantees.

---

# 🤖 Agentic Coding

Nexora-Qwen-Coder-4B is designed for workflows where the model can interact with an external environment rather than simply returning a single static answer.

A typical agent loop can be represented as:

```text
┌──────────────────┐
│   User Request   │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│ Understand Task  │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│ Plan Solution    │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│ Select Tool      │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│ Execute Action   │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│ Inspect Feedback │
└────────┬─────────┘
         │
         ▼
      Success?
       ╱    ╲
     Yes     No
      │       │
      ▼       │
┌───────────┐ │
│  Final    │ │
│  Answer   │ │
└───────────┘ │
              │
              └──────► Iterate
```

This makes the model suitable for local environments that expose tools such as:

* File operations
* Terminal execution
* Code search
* Repository inspection
* Build systems
* Test runners
* Linters
* Debugging environments

Typical agent workflows may include:

```text
Read → Plan → Act → Observe → Verify → Repair
```

Tool-call reliability depends on the application's prompt template, tool definitions, schema design, and execution environment.

---

# 🧩 Recommended Use Cases

| Use Case                  |  Fit  |
| :------------------------ | :---: |
| **Code Generation**       | ★★★★★ |
| **Debugging**             | ★★★★★ |
| **Tool Calling**          | ★★★★★ |
| **Local Coding Agents**   | ★★★★★ |
| **Code Explanation**      | ★★★★★ |
| **Refactoring**           | ★★★★☆ |
| **Repository Analysis**   | ★★★★☆ |
| **Technical Reasoning**   | ★★★★☆ |
| **Documentation**         | ★★★★☆ |
| **Software Architecture** | ★★★☆☆ |

### Best suited for

* Code generation
* Code completion
* Debugging
* Bug localization
* Error analysis
* Refactoring
* Repository inspection
* Tool calling
* Local coding agents
* Multi-step development tasks
* Technical reasoning
* Developer assistants

---

# 🖥️ Built for Local AI

The **4B parameter class** is intentionally compact.

Nexora-Qwen-Coder-4B aims to provide a practical balance between:

```text
                    CAPABILITY
                        ▲
                        │
                        │       ● Nexora-Qwen-Coder-4B
                        │
                        │
                        │
                        └────────────────────────►
                            LOCAL EFFICIENCY
```

The goal is straightforward:

> **Deliver useful coding and agentic capabilities while remaining practical to run locally.**

Potential deployment scenarios include:

* Local development
* Personal coding assistants
* Offline workflows
* Lightweight coding agents
* Developer workstations
* Experimental AI agents
* Privacy-focused development environments

---

# 📦 Model Files & Quantization

Nexora-Qwen-Coder-4B is available in **GGUF** format for efficient local inference.

| Quantization | Recommended For                                      |
| :----------- | :--------------------------------------------------- |
| **Q4_K_M**   | Best balance of quality, memory, and speed           |
| **Q8_0**     | Higher quantized quality with increased memory usage |

### Recommended: `Q4_K_M`

For most users, `Q4_K_M` provides a strong balance between:

`Quality` · `Memory` · `Speed`

### Higher Quality: `Q8_0`

Recommended when memory usage is less restrictive and higher quantized fidelity is preferred.

---

# 🚀 Quick Start

## llama.cpp

Run the model directly from Hugging Face:

```bash
llama-cli -hf guell00/Nexora-Qwen-Coder-4B --jinja
```

Start an OpenAI-compatible local server:

```bash
llama-server -hf guell00/Nexora-Qwen-Coder-4B --jinja
```

> Command availability may depend on your installed `llama.cpp` version and the model files available in the repository.

---

# 🧠 Context Length

The model was fine-tuned using sequences reaching approximately **32K tokens**.

The underlying Qwen 3.5 family may support larger context windows depending on the specific architecture and inference backend.

Long-context performance depends on:

* Backend support
* RoPE configuration
* YaRN scaling
* Quantization
* KV cache
* Available memory
* Context length

When extending beyond the training distribution, users should validate performance on their own workloads.

### Example: Extended Context with llama.cpp

```bash
./llama-server \
  -m model.gguf \
  --ctx-size 131072 \
  --rope-scaling yarn \
  --rope-scale 4 \
  --yarn-orig-ctx 32768
```

> **Important:** Increasing `--ctx-size` alone does not guarantee reliable long-context behavior.

---

# 🎯 Deterministic Coding Configuration

For highly deterministic coding, debugging, and code-repair workflows:

| Parameter              |  Value |
| :--------------------- | :----: |
| **Temperature**        |   `0`  |
| **Top P**              | `0.95` |
| **Top K**              |  `40`  |
| **Min P**              | `0.05` |
| **Repetition Penalty** |  `1.1` |
| **Presence Penalty**   |  `Off` |
| **Max Tokens**         |  `Max` |

```text
Temperature:        0
Top P:              0.95
Top K:              40
Min P:              0.05
Repetition Penalty: 1.1
Presence Penalty:   Off
Max Tokens:         Max
```

For creative programming, brainstorming, or exploratory generation, increasing the temperature may produce more diverse outputs.

For debugging and code repair, lower temperatures generally provide more deterministic results.

---

# 🔧 Built With

| Technology     | Role                               |
| :------------- | :--------------------------------- |
| **Qwen**       | Base model family                  |
| **Unsloth**    | Fine-tuning & conversion workflows |
| **GGUF**       | Efficient local model format       |
| **llama.cpp**  | Local inference                    |
| **benchlocal** | Coding & agent evaluation          |

---

# ⚠️ Limitations

Nexora-Qwen-Coder-4B is a compact **4B-class model** and should be evaluated accordingly.

It may struggle with:

* Extremely large repository-wide changes
* Complex multi-file dependencies
* Highly specialized professional domains
* Tasks requiring extensive external knowledge
* Long autonomous workflows without verification
* Deep architectural decisions involving large systems

The model should be treated as a **coding assistant**, not a fully autonomous software engineer.

Generated code should always be:

```text
REVIEWED
   ↓
TESTED
   ↓
VALIDATED
   ↓
DEPLOYED
```

Applications should verify generated code before using it in production environments.

Depending on the inference template and runtime configuration, the model may generate reasoning content inside:

```text
<think>
...
</think>
```

Applications may parse, hide, or otherwise handle these sections according to their requirements.

---

# 🙏 Acknowledgements

Special thanks to:

* The **Qwen team** for the base model family.
* The **Unsloth team** for efficient fine-tuning and conversion tooling.
* The open-source AI community for datasets, tools, and research.
* Contributors supporting local hardware testing and evaluation.

---

# 📜 License

This model is released under the **MIT License**.

Please review the licensing terms of the underlying base model and any third-party components used in your deployment.

---

# ⚖️ Disclaimer

Nexora-Qwen-Coder-4B is provided for:

**Research · Development · Experimentation · Local Inference**

Actual performance may vary depending on:

* Quantization method
* Inference backend
* Hardware
* Context length
* Prompt formatting
* Sampling parameters
* Evaluation methodology

Benchmark results represent specific evaluation runs and should not be interpreted as guaranteed performance across all environments or tasks.

Always review, test, and validate generated code before deploying it to production systems.

---

<div align="center">

# 🧠 Nexora

### Intelligence. Code. Evolve.

**Built for developers who want capable AI coding assistance running locally.**

<br>

`guell00/Nexora-Qwen-Coder-4B`

<br>

[Hugging Face](https://huggingface.co/guell00/Nexora-Qwen-Coder-4B)

</div>