Technical Documentation: gemma4-coder:12b-mlx

1. Overview

Model Identifier: mohamedbastawih/gemma4-coder:12b-mlx
Library Source: ollama.com/mohamedbastawih/gemma4-coder:12b-mlx
Role: Local OpenCode Agent
Primary Objective: To serve as the intelligence engine for an autonomous coding agent capable of interacting with local development environments, managing file systems, and executing code.

Gemma4-e4b-coding is specifically tuned to bridge the gap between a standard LLM and a functional software engineer. It does not just suggest code; it generates actionable instructions (tool calls) to manipulate files and run commands locally.

2. Agentic Capabilities

The model is optimized for the following “Agent” behaviors:

🛠️ Tool-Use & Function Calling

The model is trained to recognize when a natural language request requires a system action. Instead of a conversational response, it outputs a structured payload that a local controller (the Agent Wrapper) can execute.

📂 Local File System Integration

The model is designed to work with an agent that provides the following permissions: * Read/Analyze: The ability to ingest existing codebases to understand context before proposing changes. * Write/Modify: The ability to create new files or apply precise edits to existing source code. * Project Mapping: The ability to list directories to navigate complex project architectures.

🔄 Autonomous Coding Loop

The model operates on an Observation \(\rightarrow\) Thought \(\rightarrow\) Action cycle: 1. Observe: Read the file or error message. 2. Think: Determine the logical fix or feature implementation. 3. Act: Call the write_file or execute_cmd tool.

3. Implementation & Architecture

Model Acquisition

To use this model as the core of your agent, pull it directly from the Ollama library:

ollama pull mohamedbastawih/gemma4-e4b-coding

The Agent Workflow

Because LLMs cannot directly touch your hardware, Gemma4-e4b-coding acts as the Brain, while a local script (Python/Node.js) acts as the Hands.

Sequence of Operation: 1. User Input: “Update the API endpoint in server.js to handle 404 errors.” 2. Model Inference: Gemma4-e4b-coding triggers a tool call: \(\rightarrow\) {"action": "read_file", "path": "server.js"} 3. Local Wrapper: Reads the file \(\rightarrow\) Feeds content back to the model. 4. Model Inference: Processes code \(\rightarrow\) Generates the fix: \(\rightarrow\) {"action": "write_file", "path": "server.js", "content": "...updated code..."} 5. Local Wrapper: Commits the change to the disk.

Supported Tool Signatures

The model is optimized to produce the following JSON-style calls:

4. Usage Guide

Starting the Agent

Depending on your local wrapper, you can run the model via:

ollama run mohamedbastawih/gemma4-e4b-coding

Recommended Prompting Strategy

To get the best agentic performance, provide a system prompt that reinforces its role:

“You are the OpenCode Agent. You have access to a local filesystem. Do not just explain the code; use the provided tools to read, write, and verify your changes. Always read a file before editing it.”

5. Safety & Security

⚠️ Warning: This model is designed to generate commands that modify your local files.

• Sandboxing: It is highly recommended to run the agent inside a Docker container or a dedicated project directory.
• Confirmation Layer: We recommend implementing a “Human-in-the-Loop” verification step in your wrapper where the user must approve any write_file or execute_cmd action.
• Read-Only Mode: For auditing code, the agent can be restricted to only using the read_file and list_dir tools.

Comparison Summary