Gemma 4 26B Runs on 2GB RAM: Open-Source Engine Brings Frontier AI to Any Mac

A solo developer built TurboFieldfare, a custom Swift and Metal runtime that runs Google's 26-billion-parameter Gemma 4 model in just ~2 GB of RAM on any Apple Silicon Mac, even the 8 GB MacBook Air. The open-source project hit 767 points on Hacker News.

A solo developer has built an open-source engine that runs Google's Gemma 4 — a 26-billion-parameter model — in roughly 2 GB of RAM on any Apple Silicon Mac, including the base 8 GB MacBook Air.

The project, called TurboFieldfare, is a custom Swift and Metal runtime that keeps only the model's shared 1.35 GB core and FP16 KV cache in memory. Instead of loading the full 14.3 GB checkpoint into RAM, it streams only the expert weights needed for each token from SSD on demand.

The approach is model-specific rather than a wrapper around existing frameworks like MLX or llama.cpp. The developer, known on GitHub as drumih, wrote the entire runtime — streaming installer, CLI, native Mac app, and Metal compute kernels — in Swift 6.2 targeting macOS 26 and Metal 4.

On an 8 GB M2 MacBook Air, the engine achieves 5.1–6.3 tokens per second during decode. On a 24 GB M5 Pro, that jumps to 31–35 tok/s. The developer's experiment log summarizes 103 measured results across kernels, caching, I/O, prefill, and decode performance.

The project ships as six Swift packages: a native Mac app, a command-line interface, a loopback OpenAI-compatible Chat Completions server, and a streaming model installer that fetches and repacks the pinned Hugging Face checkpoint directly — never materializing the full source checkpoint on disk during installation.

Gemma 4 26B-A4B is Google's mixture-of-experts model with 26 billion total parameters and roughly 3.88 billion active per token. Weights use MLX affine 4-bit quantization with 8-bit router precision.

The project topped Hacker News with 767 points and drew 268 comments. Commenters noted the significance of running frontier-class models on consumer hardware without cloud dependency. Others pointed to the growing ecosystem of solo-built, model-specific runtimes — from llama.cpp to MLX to now TurboFieldfare — each pushing the boundary of what "runs locally" actually means.

The complete source is available under the MIT license.

Sources: GitHub, Hacker News discussion