One test has been making rounds in the local AI community: build a working Doom-style raycaster game in HTML and JavaScript, then see which model gets it right without melting your hardware.
For at least one developer running on a 64GB MacBook, Gemma 4 26B passed cleanly - while the previous community favorite struggled. Google's open-source model handled the raycaster task without pushing the system near its memory limits, outperforming the 4-bit quantized version of Qwen 3 Coder, which had held the top recommendation for local coding work.
Quantization is a compression process that reduces an AI model's memory footprint by storing its internal parameters with less precision. A 4-bit quantized model applies aggressive compression - useful for running larger models on limited hardware, but sometimes at the cost of reliability. In this case, the Qwen 3 Coder 4-bit variant not only taxed the Mac's memory but also showed problems with tool use - the model's ability to call external functions like code execution or web searches rather than just generating text. Gemma 4 26B handled both without issue.
Gemma 4 is Google's open-source model family, meaning the weights are freely available and run entirely on local hardware. No per-query API fees, no data leaving your machine. At 26 billion parameters, it sits in a size range that 2026 consumer hardware can handle without dedicated AI accelerator chips.
The 26B appears to be the point in Google's model family where performance and hardware demands have finally balanced out for everyday local use.