Summary
Fahd Mirza walks through the local installation and testing of NeoHorse-1-4B, a small language model from TokenRhythm built on top of Qwen 3.5. What distinguishes this model from similar parameter-count releases is its training methodology: instead of learning from static datasets, NeoHorse was trained on logs generated by a live AI routing system — real records of which model was called for which task and whether the outcome was successful. Those logs were fed back into training ordered from easy to hard, with a stronger teacher model providing live corrections in a second pass. The result is a self-improving loop where each deployed version generates new interaction logs that feed the next training run.
Mirza demonstrates the model running locally via vLLM and tests it inside a Hermes agent harness against a real-world Glacial Flood Early Warning dashboard he built with Docker, a React frontend, and a Redis-backed database. The agent is given no hints about the location or nature of a critical bug — a false all-clear that reports zero people in an inundation zone even during a simulated breach — and successfully identifies and patches all affected API endpoints through extended chain-of-thought reasoning.
The video also includes a creative writing test to probe the model’s language generation beyond code, and covers VRAM consumption figures for the vLLM serve setup. Viewers interested in running capable small models locally without expensive hardware will find practical setup guidance alongside a clear explanation of NeoHorse’s recursive self-improvement training concept.
📺 Source: Fahd Mirza · Published September 13, 2026
🏷️ Format: Tutorial Demo
