Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest
AGI Dreams Podcast – March 20, 2026
Frontier Models Escape the Data Center
In this episode:
Frontier Models Escape the Data Center — A developer ran the 397B-parameter Qwen 3.5 MoE model on a MacBook Pro at 5.7 tokens/second using only 5.5 GB RAM, leveraging Apple's unified memory architecture and SSD streaming. The project, coded by Claude Opus 4.6, shows that local inference of massive models is becoming practical with each hardware generation.
The 4-Billion-Parameter Sweet Spot — NVIDIA released Nemotron 3 Nano 4B, a hybrid Mamba-Transformer model pruned for edge deployment on Jetson and RTX hardware, while IBM's…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.