Kimi K3 is insane
Just spent the afternoon test driving Kimi K3. I see what all the fuss is about around distillation.
Lately, I've been relying primarily on Anthropic's Opus 4.7/4.8 models for most of my work. It's been getting expensive and I wanted to check out the hype around Kimi to see if it's good enough as a daily driver.
Well, yeah. I'll be damned...it is.
There's a noticeable difference in speed, but response quality is more-or-less (I didn't do a deep comparison here, this is just anecdotal) the same. Cost-wise, a few tasks in and I've noticed compared to Anthropic's API it's roughly 5-10x cheaper for roughly the same output.
I get why Anthropic had a panic about this. This proves (for me, at least) that distillation is an excellent way for open models to punch at the same weight class as frontier models. I'd even argue that Kimi listens better. I routinely have to correct Opus responses, but today, Kimi just did what I asked.
To be fair, I was working on some pretty simple stuff (building some marketing pages for Fortivibe), but Kimi picked up my existing skills and implemented exactly what I asked in my own private, custom framework (read: it wouldn't have any advantage from its training set). That's nothing to sniff at.
Just a quick note here—I don't know where this goes but it's reassuring to see that there's an escape hatch from frontier models (even if access is still via an API). I have a feeling this approach will become more common over time and perpetuate the "race to the bottom" we've been seeing in the frontier camp.
First time AI hasn't felt like an excuse for a class war.