Summary
Fahd Mirza tests Qwythos 9B v2 (also called Qutoz) — a 9-billion-parameter fine-tune of Qwen 2 with deep chain-of-thought reasoning — locally on an NVIDIA RTX A6000 GPU with 48GB of VRAM. The video opens with a technical explanation of FTPO (Fine-Tuned Preference Optimization), the surgical training technique used to eliminate the repetition loop bug that plagued v1. Rather than retraining the full model, FTPO identifies the exact token position where looping begins and nudges the model’s output preference at that single point, reducing repetitive output rates from 6.7% to zero without degrading general capability.
Mirza serves the Q4_K_M quantized version via llama.cpp — consuming just over 10GB of VRAM — and configures it with the Hermes agent framework for agentic testing. Three tasks are demonstrated: autonomous multi-bug identification and repair in a full-stack iron ore price validator application (all bugs resolved correctly); vision-language evaluation using a Mexico City Metro map photo (correct sign reading of Linea 1 and Pantitlán, but faulty route logic — an honest failure Mirza flags explicitly); and a classic hat-color prisoner reasoning puzzle, which the model solves correctly in a single clean pass with no repeated phrases, directly illustrating the FTPO fix in production.
A context ceiling issue (238k token request vs. 65k configured limit) is also encountered and resolved live by restarting llama.cpp with an increased context length. The video is a practical reference for researchers and hobbyists evaluating small open-weight reasoning models for local deployment.
📺 Source: Fahd Mirza · Published July 13, 2026
🏷️ Format: Benchmark Test
