Agent-Zero-0.5b (Ollama-Optimized Agent Model)

Agent-Zero-0.5b is a highly optimized 0.5-billion parameter agent model fine-tuned using Unsloth on Qwen 2.5. It is specifically balanced to support deep chain-of-thought reasoning and native multi-turn tool calling using structured XML schemas natively supported by modern Ollama clients and interfaces like Open WebUI.

🚀 Highlights

• Deep Reasoning: Adheres strictly to the <think>...</think> tags to organize its thought process prior to taking any action or outputting responses.
• Autonomous Tool-Use: Supports XML-based tool call invocation using <tool_call> boundaries and respects <tool_response> variables seamlessly across multi-turn interactions.
• Ultralight Efficiency: Clocking in at only 0.5B parameters, this model runs incredibly fast on low-resource machines and consumer hardware.

🛠‍️ How to run in Ollama

You can easily run Agent-Zero-0.5b in your local Ollama setup using the uploaded .gguf file and the corresponding Modelfile instructions.

1. Modelfile Configuration

Create a file named Modelfile with the following configuration:

FROM ./unsloth.Q4_K_M.gguf

# Set parameters
PARAMETER temperature 0.3
PARAMETER top_p 0.9
PARAMETER stop "<|im_end|>"
PARAMETER stop "<tool_response>"
PARAMETER stop "</tool_response>"

# System Prompt for Structured Agent Control
SYSTEM ""
You are an autonomous agent. You can think using <think> tags. You can call tools using <tool_call> tags. Wait for the <tool_response> before answering.
""

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

2. Build and Run

Run these terminal commands to compile and initialize the model:

ollama create Agent-Zero-0.5b -f Modelfile
ollama run Agent-Zero-0.5b

📊 Training Mixture

This model is a customized mixture of 8,000 highly structured samples:

• Deep Reasoning (2,000 samples from OpenO1-SFT)
• Single-turn Tool Execution (2,000 samples from AymanTarig/function-calling-v0.2-with-r1-cot)
• Multi-turn Agent Loops (2,000 samples from Glaive Function Calling v2)
• General Dialogues (2,000 samples from OpenHermes-2.5)