kimi 3

quick personal observations on kimi 3

there are no moats. kimi k3 running at a fraction of the price of claude is impressive. i wired it up to claude-code yesterday and gave it a reasonable task to refactor some code based on some consolidation that i had spelled out in a couple of paragraphs of prose. it ripped through the work in about 20 seconds. now whether this is a function of the model or the hosting (at fireworks.ai) or perhaps compute constraints at anthropic, i can’t realistically say. i would have expected claude to dither about on the changes for a minute or two.

there’s something i can’t quite put my finger on re: kimi 3 - it “feels” fresher and less encumbered then the interactions with claude do. i don’t know if that will change or if there’s just accretion and baggage that i’ve now developed with claude. i also suspect i need to give claude a dusting up with the advent of the model updates.

this dude might be channeling what i’m thinking/feeling here.

i’m less bullish on the american frontier model labs than many of my peers. i increasingly find myself more firmly in the camp of those who want high quality open weight models that can be run on a variety of providers. deepseek v4-pro was an impressive move in this space and kimi 3 is a significant jump from there. i’ll likely be poking at this for more of my personal projects.