开发者发布了一个名为TurboFieldfare的开源推理引擎，用Swift和Metal编写，能够在任何M系Mac上运行4-bit量化的Gemma 4 26B模型，内存占用仅约2GB[1]。该引擎通过在RAM中保持模型共享部分和KV缓存，同时从SSD流式传输路由专家的方式，有效解决了内存限制问题[1]。模型的4-bit量化权重占用约14GB[1]，首次运行将从Hugging Face下载15GB的权重[1]。
性能表现方面，在8GB M2 MacBook Air上，该引擎可实现5–6 tok/s的生成速度[1]；在M5 MacBook Pro上则可达到31–35 tok/s[1]。该引擎支持OpenAI兼容的本地服务器，具备流式传输和工具调用功能[1]。开发者为实现这一方案进行了超过100次实验[1]。
Developers have released TurboFieldfare, an open-source inference engine designed to run the 4-bit quantized Gemma 4 26B model on any M-series Mac while consuming approximately 2 GB of memory [1]. Built using Swift and Metal, the engine overcomes memory constraints by keeping model shared components and key-value caches in RAM while streaming routed experts from the SSD [1].
The engine delivers varying performance depending on hardware capabilities [1]. On an M2 MacBook Air with 8 GB of RAM, it achieves generation speeds of 5–6 tokens per second, while an M5 MacBook Pro reaches 31–35 tokens per second [1]. The 4-bit quantized model weights total approximately 14 GB, and users will need to download 15 GB of weights from Hugging Face during their first run [1]. The tool supports an OpenAI-compatible local server with streaming and tool-calling capabilities [1], and the developers report conducting over 100 experiments during development [1].