TurboFieldfare is a groundbreaking inference engine that runs Gemma 4 26B-A4B models on Apple Silicon Macs requiring only ~2GB of system RAM.
This comprehensive series walks through everything from architectural design principles and low-level performance tuning to practical deployment.
Series Navigation Index#
1. TurboFieldfare (Part 1): Run 26B LLM in 2GB RAM#
- Explores local LLM memory bottlenecks on limited RAM configurations.
- Unpacks the Mixture-of-Experts (MoE) design of Gemma 4 26B-A4B.
- Demonstrates how keeping a ~1.35GB shared core in RAM while streaming experts on-demand from NVMe SSD bypasses traditional memory constraints.
2. TurboFieldfare (Part 2): Swift & Metal Runtime Optimization#
- Deep dive into the custom runtime built natively with Swift 6.2 and Metal Compute Shaders.
- Examines Least Frequently Used (LFU) caching for hot expert layers.
- Reviews Chunked Prefill and weight quantization techniques used to mitigate SSD read latency.
3. TurboFieldfare (Part 3): CLI, Mac App & OpenAI Server#
- Setup requirements and build steps for macOS 26+ and Xcode 26.
- Explains CLI usage, native SwiftUI app features, and local OpenAI-compatible API server configuration.
- Provides hardware performance benchmarks across Apple M2 Air and M5 Pro chips.
[!TIP] Inspect full source code and over 100 benchmark experiments at the official repository: drumih/turbo-fieldfare ↗.
References#
- Official Repository: drumih/turbo-fieldfare ↗
- Technical documentation on MoE architectures and Apple Silicon I/O streaming.