[mindwok]

Does anyone else feel like the writing is on the wall for a future of local models? Spamming data centres everywhere, powering them, having to commit insane capital to hardware, all the effort to serve inference over a network reliably - when here we are with a frontier model nearly running on a laptop.

Local AI on your device seems like a much more likely future to me than datacenters in space. For inference at least, training is another story.

[nomel]

0.01 tk/s on an M1 Max is not "nearly". This is completely unusable, and in no way cost effective.

0.01 tokens per second means 1 million tokens ($3 worth of API usage [1]) takes 3.2 YEARS.

[1] https://www.kimi.com/resources/kimi-k3-pricing

[mindwok]

Ok in terms of running a 2.8T parameter model, that's true.

Looking more broadly though, a model I can run on my laptop (Gemma 4) is ~4 points away from GPT-5.3 codex or Sonnet 4.5 on arena.ai LLM leaderboard. Those models were SOTA less than a year ago.