Kimi K3 is really, really good.

Theo please why don’t you make short videos? 40-50 Mins is too long this could be summarized in 5-10 Mins video tbh.

Great breakdown. For those wondering what 'Open weight models are', saved you a search.

theo it's 6am (i'm just optimizing my reset windows don't worry)

All I see is this when I hear Kimi ugh

those caveats about china reading your code makes no sense. Who have been manipulating data and personal info of US citizens are US based companies.

One note: Kimi's sub has 5h, weekly, and monthly limits - but it's hard to realize how it works. The numbers I'm seeing now suggest I'll barely hit the monthly cap just by coding. However, I've already hit the monthly limit before when I was also using the in-app swarm.

Kimi K3 lives up to the hype. Strong performance and another real step forward in the model landscape. It is live on CreateOS now

I honestly wonder how long the Chinese government allows this to progress before they start pulling the same national security moves the US is. The idea that this will really have open weights in a short time is… unbelievable. I hope to be wrong.

Even if I don't necessarily plan on using it, the competition is great.

Hey @grok, how can I watch this 40 minutes video in like 2?

Kimi K3 is really, really good. I completely agree with you, especially after trying it myself. Here's the demo:

I get your point of local models for us but think about the enterprise. Walmart or JPMC can easily put up hardware worth millions to host these and keep their data private and not have the contracts with OpenAI or Anthropic. It would be a big blow

Woah this is pretty interesting

Agreed. The date I'd circle is July 27, when the weights actually drop. Good is temporary when it's someone else's endpoint.

They really need to partner with a US provider like cerebras

Kimi was so good that Anthropic decided to shut down Fable 5 earlier than expected

loved the kimi k3 model. the architecture of kimi k3 is amazing, i studied the architectural diagram properly and explained how a token enters and exits the blackbox in very simple words.

ChatGPT moment for open weights.

Yes it is !!!

Kimi is good! But Hungry as f....

I've been seeing all good reviews

I also heard that many people praise it, but I also heard criticism about the number of tokens, which is quickly spent at the moment

gotta check out what they did with the context window

huh, that fast

Cool we can try it out

It hasn't even been tested thoroughly. All we know is that it's good at 3D and UI. It's quite expensive, and Kimi models aren't known for token efficiency, so more tests are needed. Overall, it's a good model, but Codex and Claude Code subscriptions are probably still cheaper.

can you do a video on how to set up different models on claude code. I watched your last video but like you spent basically no time talking about how to put 5.6 sol in claude code??? I think a step by step guide would be nice thanks!

one of the most impressive releases this year

ok but do you want to expose yourself to a potential data-exfiltration mechanism

one more contender pushing for the top spot

Heading straight away to test it and see how "really really good" it is

extremely bullish on Cursor Grok 5

nice....was waiting for this video