Author(s): Caspar Bannink

Originally published on Towards AI.

The Model That Matches GPT 4.1, Is Private and Free

I have been waiting for a 12B model that can replace what I am paying OpenAI for, and Gemma 4 12B is the first one this year that actually does. I pulled it last Tuesday, ran it on a quantization I can fit on my dev box, and the first thing I did was put it against a battery of agentic coding tasks I would normally hand to GPT-4.1. It did not win every one. It won enough of them that I have not opened ChatGPT for routine code work in six days. That is the headline. The rest of this is the receipts.

After introducing Gemma 4 12B, the author argues that its “agentic coding” capability—not just raw benchmark specs or model-sheet promises—is what made it convincing in real local workflows. They cite first-party evaluation numbers (including Tau2-bench, LiveCodeBench, and Codeforces ELO) and highlight native function-calling support and an official local example via llama.cpp that demonstrates building an image-processing Gradio app. They also describe hardware feasibility and why the encoder-free multimodal architecture (plus reduced memory footprint) and native audio input make the model practical on a developer’s existing machine, enabling screen/image/audio-aware agent loops without separate vision APIs. Finally, they note important limitations (frontier-API gap, not being a true frontier coder, bounded long-horizon autonomy, and that the 12B is the sweet spot rather than the largest variant), then explain what they plan next: keep the model loaded, integrate it into a screen-aware assistant, and test its audio-native input on customer-support recordings, using their own usage to measure how well marketing matches reality.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI