Top AI Product

A 26B model in 2GB of RAM sounds impossible. That’s exactly why it hit 556 points on Hacker News and 865 GitHub stars in a day.

What it actually is

TurboFieldfare is a local inference engine written in Swift and Metal that runs Google’s Gemma 4 26B-A4B — 14.3GB of weights — on an 8GB M-series Mac. It ships as a Swift library, a CLI, an OpenAI-compatible loopback server, and a native SwiftUI app. Not a hosted API. You run it on your own machine, offline.

The trick is a MoE cheat: only the 1.35GB shared core and KV cache stay resident. The rest of the expert weights stream off the SSD per token, backed by a 16-slot LFU cache. So a big model behaves like a small one, and the disk does the heavy lifting.

Why it’s worth watching

The numbers hold up. An 8GB M2 MacBook Air pushes 5.1–6.3 tokens/s; a 24GB M5 Pro hits 31–35. That’s a frontier MoE model running usably on a laptop most people already own — a clean proof that local big-model inference isn’t gated on 64GB machines anymore.

You Might Also Like

• Google Ships Gemma 4 Multi Token Prediction Drafters 2 7 3 5x Faster Inference Free
• Google Gemini 3 5 pro ga Ships With a 2m Token Context Window the Biggest in any Production Model
• 397 Billion Parameters on a 48gb Macbook Flash moe Turns Apples 2023 Research Into Reality
• Google Turboquant Squeezes llm Cache to 3 Bits 6x Less Memory 8x Faster Zero Accuracy Loss
• 708 Github Stars in 48 Hours Claude Token Efficient Universal Claude md and the Fight Over Claudes Most Expensive Habit