Open-source engine runs Gemma 4 26B in ~2 GB RAM on M-series Macs
Original: Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac
Why This Matters
Demonstrates that frontier-scale MoE models can run on consumer hardware with constrained RAM via SSD streaming.
Developer drumih released TurboFieldfare, an open-source Swift + Metal runtime that runs Google's Gemma 4 26B-A4B model in approximately 2 GB of RAM on any Apple Silicon Mac, including 8 GB models, by streaming model experts from SSD on demand.
TurboFieldfare is an open-source inference engine written in Swift and Metal, designed to run Google's instruction-tuned Gemma 4 26B-A4B large language model within roughly 2 GB of RAM on any M-series MacBook — including entry-level 8 GB configurations. The model's full weights total 14.3 GB, which would normally far exceed the available memory on base-tier Macs. TurboFieldfare addresses this by keeping only the shared 1.35 GB core and an FP16 KV cache resident in memory, then streaming only the specific Mixture-of-Experts (MoE) expert layers required for each token directly from SSD storage. This selective loading approach avoids materializing the entire model in RAM at any point. The project is hosted on GitHub under the Apache-2.0 license and has accumulated over 574 stars since its initial commits. The repository includes documentation covering quick-start instructions, local server setup, benchmarks, and details on the underlying technical approach. Contributions and benchmark results from the community are also solicited.