About Unified Memory

How fast is local AI on a Mac, really, and which model should you run? Unified Memory measures it.

We benchmark open models and AI coding agents on Apple Silicon, and publish a short daily roundup of AI and tech news. Every benchmark number comes from a run on our own M4 Max with 36 GB of unified memory, with the exact settings published next to the result.

How we benchmark

• We write our own test sets for real jobs people do, instead of reusing public leaderboards a model may have trained on.
• Each item is scored by code where the answer can be checked exactly, and by a blind panel of three AI judges (median taken) where it can't.
• Speed and memory are measured on the same machine, one model at a time, and reported next to quality rather than blended into it.
• Every scored item is on this site: the prompt, what we expected, what each model answered and how it was scored.

How the news is made

AI agents scout the day's stories, research them from primary sources and draft the script. A different AI model then fact-checks every claim against its primary source; findings are fixed and re-checked for up to two rounds, and any story that still fails is dropped before anything is narrated. A human can veto any story before it is written. The narrator voice is synthetic. Each edition page here lists every claim with its source and the fact-check log.

Contact

Through the channel's About page on YouTube, or leave a comment on any video.