BREAKING
397B Model Runs on a MacBook Pro
0B
total params
0B
active params
0GB
unified memory
How SSD Expert Streaming Works
1mmap 5.5GB non-experts
2Stream experts from SSD
3OS page cache holds 42GB
4FMA Metal kernel runs
Speed by Quantization Scheme
2-bit5.74
4-bit + FMA4.36
4-bit base3.9
Promise vs PoC Limits
UpsideOn-device
Tool calling on a laptop
Privacy, low latency
iPhone 17 Pro fork appeared
LimitsPoC
2-bit breaks JSON
Heat and throttling risk
Heavy SSD dependence
Built in 24 Hours, 90+ Experiments
AI NEWS BLITZ
A new engine called Flash-MoE runs a 397 billion parameter model on a laptop.