Summary
Fahd Mirza covers the most significant update to Google’s Gemma 4 model to date: 11 targeted fixes spanning tool calling, chat formatting, thinking content retention, and vision — all bundled into the same model weights without a name change. The video focuses specifically on the 4-bit Quantization-Aware Training (QAT) release, which compresses the 12-billion-parameter model from roughly 24 GB to 6.72 GB on disk, making it practical to run on a single consumer or prosumer GPU.
The tutorial explains QAT clearly: unlike post-training quantization, QAT simulates 4-bit rounding during training so the model learns to compensate for precision loss before it is applied. The result, according to Google, is quality that stays close to the full-precision model. The video also covers Multi-Token Prediction (MTP) drafter heads baked into the weights — a speculative decoding approach that delivers roughly 20% more tokens per second using a compact 254 MB auxiliary file, with no separate model download required.
The real-world test runs Gemma 4 via llama.cpp with the Hermes agent framework against a FastAPI/SQLite port freight tracking application. The agent is given only a high-level goal — find and fix all bugs, start the server, prove it works — with no file paths or hints. The safety-check bug involving draft depth validation is identified and resolved autonomously, demonstrating that Gemma 4’s tool-calling fixes translate to reliable multi-turn agentic behavior at low VRAM cost.
📺 Source: Fahd Mirza · Published July 18, 2026
🏷️ Format: Hands On Build
