Open-Source Engine Runs Gemma 4 26B on M-Series Macs with 2GB RAM

gitpusher42· July 29, 2026 View original

Summary

An open-source inference engine, TurboFieldfare, allows running the 4-bit quantized Gemma 4 26B model on M-series Macs using only 2GB of RAM by streaming model experts from SSD. It achieves 5-6 tokens/second on an M2 MacBook Air and 31-35 tokens/second on an M5 MacBook Pro, also offering an OpenAI-compatible local server.

A new open-source inference engine, named TurboFieldfare, has been developed to enable the execution of the Gemma 4 26B large language model on Apple's M-series Macs, remarkably utilizing only about 2GB of RAM. This is achieved by strategically keeping the model's shared components and KV cache in RAM while streaming necessary "experts" from the SSD, overcoming the challenge of the model's 14GB quantized weight size. The engine, written in Swift and Metal, demonstrates impressive performance, ranging from 5-6 tokens per second on an 8GB M2 MacBook Air to 31-35 tokens per second on an M5 MacBook Pro. Additionally, it includes an experimental OpenAI-compatible local server, supporting streaming and tool calls.

Why it matters

Developers and researchers can now run powerful large language models locally on consumer-grade M-series Macs, significantly reducing reliance on cloud resources and enabling privacy-preserving AI applications.

How to implement this in your domain

  1. 1Download and install the TurboFieldfare Mac application on an M-series Mac.
  2. 2Experiment with running the Gemma 4 26B model locally for various NLP tasks.
  3. 3Integrate the OpenAI-compatible local server into existing development workflows for testing.
  4. 4Contribute to the open-source project by providing feedback or code enhancements.
  5. 5Explore possibilities for privacy-focused AI applications leveraging on-device inference.

Who benefits

Software DevelopmentAI ResearchEducationCybersecurityCreative Arts

Key takeaways

  • TurboFieldfare enables large language models to run on M-series Macs with minimal RAM.
  • The engine uses an innovative SSD streaming technique for model experts.
  • It offers an OpenAI-compatible local server for development and testing.
  • On-device AI reduces cloud dependency and enhances privacy.

Original post by gitpusher42

"Hi HN, I built a specialized inference engine for running 4-bit Gemma 4 26B-A4B-IT on any M-series Mac using about 2 GB of RAM. It is called TurboFieldfare and is written in Swift and Metal. I have always adored on-device AI. It feels like magic that you can run a powerful NN on…"

View on X

Originally posted by gitpusher42 on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses