README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - crewai
  - code-generation
  - multi-agent
  - qwen3
  - gguf
  - unsloth
  - thinking
---

# CrewAI Code Generation - Qwen3 8B (GGUF)

Fine-tuned Qwen3-8B models for generating CrewAI multi-agent code from natural language descriptions.

## Models Available

### V2 Thinking Models (Recommended)
| File | Size | Description |
|------|------|-------------|
| crewai-qwen3-8b-thinking-v2-q4_k_m.gguf | 4.7 GB | 4-bit quantized, best balance |
| crewai-qwen3-8b-thinking-v2-q8_0.gguf | 8.2 GB | 8-bit quantized, higher quality |

**V2 Improvements:**
- Native Qwen3 chat template (ChatML format)
- Higher LoRA rank (r=32 vs r=16)
- Thinking mode with reasoning tags
- Better structured outputs

### V1 Models (Legacy)
| File | Size | Description |
|------|------|-------------|
| qwen3-8b.Q4_K_M.gguf | ~4.7 GB | 4-bit quantized |
| qwen3-8b.Q5_K_M.gguf | ~5.5 GB | 5-bit quantized |
| qwen3-8b.Q8_0.gguf | ~8.2 GB | 8-bit quantized |

## Usage with Ollama

1. Download the GGUF file
2. Create a Modelfile:

```
FROM ./crewai-qwen3-8b-thinking-v2-q4_k_m.gguf

TEMPLATE """{{- if .System }}
<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

SYSTEM """You are a CrewAI code generation expert. When given a task description and required inputs, generate complete, working CrewAI Python code that includes:
- All necessary imports (crewai, crewai_tools)
- Agent definitions with roles, goals, backstories, and tools
- Task definitions with proper context dependencies
- Crew instantiation with appropriate process type
- Kickoff code with the provided inputs

Think through the problem step by step before generating code."""

PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 8192
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"
```

3. Create and run:
```bash
ollama create crewai-qwen3-8b -f Modelfile
ollama run crewai-qwen3-8b
```

## Example Prompt

```
Create a CrewAI crew for analyzing competitor websites.
Required inputs:
- competitor_urls: list of URLs to analyze
- analysis_focus: what aspects to focus on
```

## Training Details

- **Base Model**: Qwen/Qwen3-8B
- **Fine-tuning**: Unsloth + QLoRA (r=32, alpha=32)
- **Dataset**: 2,500 CrewAI code examples
- **Training**: 3 epochs on RTX 4090

## License

Apache 2.0 (same as base Qwen3 model)