Moonshot Drops Kimi K2.6 – Cheap, Powerful, and Worth Trying for Builders

+

Moonshot just dropped a new, more powerful Kimi: K2.6. I tested it this morning, and it’s the best open source model I’ve used yet.

Moonshot claims that Kimi outperforms GPT-5.4 and Claude Opus 4.6 on some benchmarks. Better yet, Kimi is dirt cheap to run, making it a popular choice for developers.

I ran Kimi through 3 difficult tests this morning: researching startup funding rounds, prepping me for a meeting, and helping me study Japanese. Overall, Kimi turned in a strong performance, although it struggled with some web search tasks.

Let’s see what the new Kimi can do!

Round #1: Researching the Cursor Seed Round

SpaceX recently agreed to acquire Cursor for $60 billion. Almost every seed investor in the world is slapping their forehead, wishing they’d been in on it… including me!

I’m curious: what did the company look like at the time? Maybe studying such a successful company can help me find the next Cursor.

Cursor had already passed $1 million ARR by the seed round…unusually strong traction.

Cursor also had an exceptional founder. CEO Michael Truell had interned at Google in high school, which is highly unusual.

The deal wasn’t cheap — around $60 million. But especially in 2023, seed stage companies with over $1M in ARR were quite rare.

Cursor was a hot startup with serious traction and a great founder. It had every factor I look for.

The price was steep. But that shows the importance of focusing on median entry price, rather than an absolute price limit. Expensive deals make sense occasionally, so long as they’re balanced out by cheaper ones.

Kimi did a great job, citing reliable sources and giving me the info I wanted. I’m giving this round an A!

Round #2: Stalking a Founder

Can Kimi can help me find the next Michael Truell? Next week, I’m meeting a fellow who might fit the bill…

This young man has a very interesting startup in robotics. I asked Kimi to tell me about his background and any exceptional achievements.

With a little help, Kimi pulled some interesting info (name redacted).

He managed to work in ROTC and run a startup at the same time. That’s impressive! I’m looking forward to meeting him.

Kimi found useful info, but only with a lot of coaxing. Kimi claimed it couldn’t access Linkedin, so I had to screenshot his LinkedIn and upload it.

I gave the same prompt to Grok, and it found a bunch of information on the founder right away.

I’m giving Kimi a C for this round.

Round #3: Kimi the Language Learning Coach

It’s less than a week until I ship off to Tokyo! I’m desperately trying to brush up on my Japanese. Can Kimi help?

Kimi gave great advice. It suggested I listen to the Nihongo con Teppei podcast and repeat everything he’s saying.

I pulled up a short video of Teppei’s and tried it. It was a wonderful way to get myself speaking and thinking in Japanese.

The only thing I didn’t like about Kimi’s response was that it didn’t give me a link to the podcast. I still had to Google it.

Still, Kimi was super helpful! I’ll give this round an A-

Wrap-Up

I’m giving Kimi K2.6 a B- overall in my testing. It’s light years ahead of K2.5, which I gave a D in February.

Kimi is good at analyzing information, but sometimes struggles in web searches. The hosted version of Kimi definitely isn’t the best model around.

But for a cheap, open-source model to use in your application, Kimi is a great choice.

A lot of startups begin with a frontier model like Claude or Grok and get the output they want. Then, they use a cheaper model and try to get the same results at scale.

Kimi is an excellent choice for that cheaper model. If you’re building an AI, give the new Kimi K2.6 a shot!

Moonshot Drops Kimi K2.6 – Cheap, Powerful, and Worth Trying for Builders

Share this:

Leave a comment Cancel reply