TurboFieldfare runs Gemma 4 26B MoE model in 2GB RAM on any Mac

TurboFieldfare is an open-source Swift and Metal runtime that runs Google's 26-billion-parameter Gemma 4 26B-A4B model on any Apple Silicon Mac, including an 8 GB M2 MacBook Air. Rather than wrapping MLX or llama.cpp, it's a model-specific engine built to exploit the model's mixture-of-experts architecture: instead of loading all 14.3 GB of weights into memory, it keeps only a shared 1.35 GB core and an FP16 KV cache resident, streaming the experts needed for each token from SSD. That keeps total memory use around 2 GB while only ~3.88B of the 26B parameters are active per token.

The project ships a native Mac app, a CLI, and an experimental OpenAI-compatible loopback server, all sharing the same .gturbo model directory. Its streaming installer repacks Hugging Face byte ranges directly into that format without ever materializing a second full checkpoint on disk. Measured throughput lands at 5-6 tokens/sec on an 8 GB M2 MacBook Air and 31-35 tokens/sec on a 24 GB M5 Pro. For engineers, it's a concrete demonstration of making large MoE models practical on memory-constrained consumer hardware through selective expert streaming.