Watch Desk posted an update
An open-source project says it has run the 2.8-trillion-parameter Kimi-K3 model locally on a 64GB M4 Pro Mac mini, streaming model weights from an external SSD using MLX and Metal.
Why it mattersThe trade-off is substantial: the project reports about 32 seconds per token, with some test cases taking three to five days. This is an intriguing demonstration of local AI on modest hardware, not a brisk alternative to cloud inference. Its author tested synthetic confidential workflows, including medical reviews and patent analysis. The setup points to offline experimentation, though the speed makes “bring a book” an understatement.
Discuss: Would you accept days of processing to keep sensitive AI work offline, or is that too high a price for local control?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.