Launched this week
ReteAi
Your own AI. On your Mac. No subscription, ever
8 followers
Your own AI. On your Mac. No subscription, ever
8 followers
Mesh your Apple devices into a local AI supercomputer. Rete runs models like Llama and Mistral on your Mac — and pools compute from your iPhone, iPad, or friends' Macs to handle larger models. Fully private: conversations stay on your network. Works offline. No accounts, no subscriptions, no data harvesting. One payment of $20, yours forever. Stop renting AI and start owning it.
jared.so
The "mesh your Apple devices" angle is what differentiates this from every other local-LLM wrapper — pooling compute is the missing infrastructure piece. Curious what the latency looks like once you bring in a non-LAN device (a friend's Mac across the internet): is throughput still usable, or does it degrade fast enough that the mesh ends up being effectively same-network only in practice?
@mcarmonas Yes, it degrades your math instinct is right. LAN is near-free (~2s per response with local Metal), but the WAN path through our relay hits the per-token RTT wall hard. We see ~5s/token for short responses cross-internet, exactly the compounding you'd predict.
So WAN mesh isn't really for "make my Mac faster." The use case it's actually for is memory pooling: a 70B+ model that won't fit on one device becomes runnable when a friend holds half the layers, and at that point 5s/token beats "can't run at all." For models that fit locally, LAN is where the throughput story lands.
So in practice: same-network for chat throughput, cross-internet for big-model access. Two different value props on one substrate.