26B params, 2GB RAM: your 8GB Mac just got delusions of grandeur
What happened
An open-source project called TurboFieldfare dropped a custom Swift + Metal runtime that runs Google's Gemma 4 26B-A4B model on Apple Silicon Macs using only about 2 GB of RAM. Instead of loading the full 14.3 GB model, it keeps a small shared core plus KV cache in memory and streams the needed 'experts' from SSD per token, meaning even base 8 GB MacBooks can run a 26-billion-parameter model.
Why this matters
Memory has become the real bottleneck in local AI, not compute — cloud GPUs are pricey, but so is RAM-maxed hardware. If mixture-of-experts models can stream from SSD instead of hogging RAM, suddenly 'big model, small laptop' stops being a contradiction and starts being a Tuesday.
The slightly cynical read
'Runs on 8GB Macs' is the new 'runs on a Raspberry Pi' flex — impressive, but token-per-second speed and quality often get quietly left out of the headline. Streaming experts from SSD sounds elegant until your fans spin up like a jet engine and your disk starts aging in dog years.
What to watch next
Watch for real-world benchmarks on tokens/second and whether this approach generalizes beyond Gemma's MoE architecture to other open models. If it holds up, expect a wave of copycat 'run giant models on your grandma's MacBook' projects within weeks.