What's this about?
Two well-known engineers tried to run the open-weight Kimi K3 on a single high-end laptop. David Ha (hardmaru) got it going on an M5 Max at 0.3 tokens per second — glacially slow — and antirez streamed the official 1.6TB of compressed weights through a 128GB M5 Max as a stress test.
What exactly does "0.3 tokens per second" mean?
A "token" is roughly a word-piece; models generate text one token at a time. Fast chat models do dozens to hundreds per second. 0.3 tok/s means about one word every three seconds — you'd wait minutes for a single paragraph. The bottleneck is size: K3 is a 2.8-trillion-parameter model, and even heavily compressed (mxfp4) its weights are ~1.6TB, far more than any laptop's memory, so most of it streams from disk.
Why is this worth noting?
It's the reality check under yesterday's "first open-weight frontier model" excitement. "Open weights" means you *can* download and run it — but "frontier-size" means practically nobody runs it at home yet. hardmaru's dry "probably as slow as it will ever be" is the optimistic read: hardware and quantization will close the gap over time.