Faster chips, new bottleneck 🤯 Two clustered M5 Ultra Macs ran SLOWER than one: 35 tok/s vs 53. Which is strange, considering my experiments with M3 Ultra were the opposite. Keep the GPUs awake and they jump to 58 tok/s.

then...I ended up running tensor parallel, but deeepseek v4 flash has only one KV head, so the trick is to copy it to both machines, and split all the other layets. The result: the best prefill of any setup (1,810 tok/s, +23% over one Mac) but the worst decode (32 tok/s). Every token needs 60+ syncs between the Macs.

Vibe coding has made building apps ridiculously fast, BUT getting that app deployed is still way more annoying than it should be. Hostman is built around making that part faster, especially if you’re already building with GitHub, Cursor, Claude, etc. I took a look at how it works 👇 (Sponsored by Hostman.)

They also have managed databases like PostgreSQL, Redis, MySQL and ClickHouse. This is the kind of stuff I usually don’t want to spend time setting up when I’m just trying to get a project running.