This 4B Local AI Scores 96.9% on Coding! (NeoHorse Agent Setup)

Neo Horse AI: A Self-Improving Local Agent for Secure Code Generation

This video introduces Neo Horse AI, a 4-billion-parameter language model designed for offline, self-improving code generation. Unlike cloud-dependent agents, Neo Horse runs entirely on local hardware, using a guided training harness to learn from its own mistakes. The creator demonstrates its installation, architecture, and a practical task—filtering sales data and generating a structured JSON report—to showcase its autonomous debugging and tool-calling capabilities.

🧠 Architecture & Self-Improvement Loop

Neo Horse employs a four-step internal routing workflow:

1. Receive complex coding instructions and process all required logic.
2. Prepare active working memory by injecting perfect execution examples into the live context.
3. Activate an automated system prompt that double-checks tool parameters, preventing formatting errors before execution.
4. Parse structured data to extract instructions clearly.

The core innovation is self-improvement via trajectory signals: every log is converted into training data, enabling the agent to avoid repeating past mistakes. A pre-dispatch audit catches JSON formatting errors in real time, ensuring clean tool calls without manual fixes.

📊 Performance & Benchmarks

• HumanEval coding benchmark: 97% accuracy, outperforming its base model (Qwen 2.5 3B).
• 10 primary evaluation suites: 65% average score—a 6% gain over the base model and competitors (e.g., Qwen 2.5 scores 59%).
• Context window: 250,000 tokens (expandable to 1 million), allowing full project files and documentation to be ingested.
• Hardware requirements: With 4-bit quantization, less than 4 GB VRAM is needed, making it accessible on entry-level GPUs.

⚙️ Practical Demonstration

The creator registers Neo Horse via Ollama, sets a 16k context window, and provides a task: filter invalid sales entries (price ≤ 0), compute total revenue and average price, and output JSON. The model generates Python code, performs a self-correction audit (identifying two invalid rows), validates calculations, and returns a clean report. The entire pipeline runs locally, with no data leaving the system.

⚠️ Limitations

• Knowledge base is limited due to 4B parameters—cannot match frontier models like GPT-4.
• Text-only inference—no visual data processing.
• Requires a local environment (e.g., Ollama) and initial configuration.

Key Takeaway

Neo Horse AI offers a reliable, secure offline alternative for developers who need automated code generation, debugging, and structured reporting without cloud dependency. Its self-improving loop and pre-dispatch auditing make it particularly suited for backend scripting (Python, C++) and tool-calling workflows where data privacy and consistent JSON output are critical.