Descriptions:
This video from Fahd Mirza walks through running Google’s Gemma 4 E4B model entirely locally using Ollama and OpenClaw — at no cost beyond compute. The E4B is an edge-optimized variant with 8 billion total parameters but an effective 4 billion parameter footprint at inference time, made possible by Google’s per-layer embedding architecture: each transformer layer gets its own small lookup table per token, enabling the model to run with the speed and memory profile of a 4B model while retaining the knowledge capacity of a much larger one. Mirza explains this clearly using a book-and-index analogy before diving into the installation.
The setup flow covers installing Ollama, pulling the Gemma 4 E4B model from Ollama’s library, and configuring OpenClaw — an open-source agentic platform — to route requests to the local Ollama endpoint. VRAM consumption in active agentic context runs above 15GB including KV cache, relevant for anyone planning hardware. The video shows both the OpenClaw terminal UI and dashboard interface once everything is connected.
For the practical test, Mirza gives the model a complex existing ant colony simulation (originally generated by Gemma 4’s 31B model) and asks it to make surgical additions: a speed control slider, a manual day/night toggle button, a population cap increase to 500, and a live population graph — all without breaking the running simulation. The test provides a realistic view of how the E4B handles tool-use, code comprehension, and file editing in an agentic loop on locally hosted hardware.
📺 Source: Fahd Mirza · Published April 03, 2026
🏷️ Format: Tutorial Demo
