CoolFace
Modelpublic

rchow93/crewai-qwen3-4b-gguf

sourceHugging Faceupdated 9mo agoView on Hugging Face
0likes40downloads
Model Card

CrewAI Qwen3-4B GGUF Models

Fine-tuned Qwen3-4B models for CrewAI code generation.

Available Models

V2 Thinking Models (Recommended)

These models use Qwen3 native chat template with <think> reasoning tags.

FileSizeDescription
crewai-qwen3-4b-thinking-v2-q4km.gguf2.4 GBBest balance of size/quality
crewai-qwen3-4b-thinking-v2-q8_0.gguf4.0 GBHigher quality

V1 Models (Legacy)

Original models using Alpaca-style template.

FileSize
qwen3-4b.Q4KM.gguf2.4 GB
qwen3-4b.Q5KM.gguf2.7 GB
qwen3-4b.Q8_0.gguf4.0 GB

V2 Improvements

  • —Native Qwen3 chat template (better tokenization)
  • —Thinking mode with <think> tags for reasoning
  • —LoRA r=32, alpha=32 (higher capacity)
  • —Trained on 2,500 CrewAI examples

Usage with Ollama

bash
# Download the model
wget https://huggingface.co/rchow93/crewai-qwen3-4b-gguf/resolve/main/crewai-qwen3-4b-thinking-v2-q4_k_m.gguf

# Create Modelfile
cat > Modelfile << EOF
FROM ./crewai-qwen3-4b-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."""

PARAMETER temperature 0.7
PARAMETER num_ctx 8192
PARAMETER stop "<|im_end|>"
EOF

# Create and run
ollama create crewai-qwen3-4b -f Modelfile
ollama run crewai-qwen3-4b

Training Details

  • —Base model: unsloth/Qwen3-4B
  • —Dataset: 2,500 CrewAI code examples
  • —Method: QLoRA with Unsloth
  • —Epochs: 3