Kimi K3.1 Pro: China's New AI, Explained Simply

Have you heard of Kimi? Most people haven’t, which is odd, because its newest model just finished a hair behind Anthropic’s best Claude on one of the toughest practical tests around.

Kimi K3.1 Pro comes from Moonshot AI, a Beijing startup, and was announced on October 23, 2026. It’s an upgraded version of Kimi K3, which came out in July. Moonshot says it’s better at coding, building 3D models, financial research and multi-step tasks, and its API costs about a third less than K3’s did. The company also says it will publish the model’s full “weights” (the trained brain itself, free for anyone to download) by November 4. You can already chat with it in Chatday.

Inside Kimi K3.1 Pro: a giant model that only wakes up a slice

Think of K3.1 Pro as K3 after a long stretch of extra coaching. Moonshot didn’t build a new brain from scratch. It took the K3 model and spent much more computing power on “post-training”, the stage where a model learns to follow instructions, reason step by step and finish tasks rather than just predict text.

The model is huge. Moonshot puts it at about 2.8 trillion parameters in total (parameters are the internal dials a model adjusts while learning). It’s also built so that only around 104 billion of them switch on for any one request. Picture a giant hospital where each patient sees only the few specialists they need. That’s how a model this big stays fast and affordable to run.

There’s also a sibling, Kimi K3.1 Pro Highspeed, which Moonshot says answers two to three times faster with the same abilities. For most people the regular version is the one that matters.

One quirk is worth knowing. K3.1 Pro always “thinks” before it answers, and you can’t switch that off. You’ll often wait a few extra seconds, but on tricky questions that pause is where the quality comes from. If you’ve wondered why some AIs pause before replying, we explained it in why your AI now thinks before it answers.

How close is it to Claude, really?

This is the claim behind the headlines, so it’s worth reading carefully.

On Terminal-Bench 4.0, a test where a model has to complete real jobs inside a computer’s command line (installing things, fixing broken setups, running scripts), Moonshot reports that K3.1 Pro scored 65.7%. Claude Opus 5.5 scored 66.4%. In plain terms, that’s a near tie with one of the strongest models available.

Two things keep that in perspective:

• It’s Moonshot’s own scorecard. Tech outlets that covered the launch pointed out the results are self-reported and not yet independently checked. An early independent ranking reportedly put it just behind Anthropic’s top Claude models, which is impressive, but behind is still behind.
• It isn’t strongest everywhere. Moonshot’s own results show it trailing the top closed models on cybersecurity tasks, and that gap is on purpose (more on that below).

Where it reportedly shines is practical, structured work. Moonshot highlights generating 3D models from a description and complex financial analysis, where it says K3.1 Pro beat the leading closed models on one finance test. If your questions look like “build me a spreadsheet model for this business plan” or “write a script that cleans up these files”, it’s worth a try.

Moonshot weakened its own model on purpose

This is the detail most coverage led with, and it’s unusual.

While training K3.1 Pro, Moonshot noticed the model was getting better and better at finding and exploiting security holes in software. That’s a problem for a model you’re about to give away to anyone, so the company says it removed hacking-related material from the training data while keeping ordinary coding practice in.

The numbers Moonshot shared, from its own internal test of exploiting real software flaws:

So the released model is still far more capable here than its predecessor. The cuts slowed it down rather than stopped it. Moonshot’s stated aim is to avoid releasing downloadable weights that push the frontier of what hackers can do. For an everyday user this changes nothing. For the wider debate about open versus closed AI, it’s an interesting sign that open-weight labs are starting to hold things back themselves.

Kimi K3.1 Pro vs Claude and DeepSeek: which to use for what

Here’s a practical guide to where K3.1 Pro fits next to other models you can open in Chatday today:

The last row matters most. No single AI wins at everything anymore, and Kimi is a good reminder that strong models now come from labs most people have never heard of. We covered that shift in the Chinese AI models catching up to ChatGPT. K3.1 Pro is the clearest example yet.

If you’ve used the earlier version, you can still find it on the Kimi K3 model page. It’s a fair way to see how much the extra training changed things.

Where to use Kimi K3.1 Pro today

Moonshot offers it in its own Kimi app and website, on its desktop and coding tools, and through its API for developers. Kimi’s free plan includes it with “limited usage”, and the company hasn’t said exactly how limited.

The other option is to open it in Chatday, next to Claude, GPT, Gemini, Grok and DeepSeek. That’s the most useful way to judge a new model, because you can ask K3.1 Pro something hard and then send the same question to Claude in the same conversation without copying anything between apps.

Here’s a good first test. Give it a messy task you’d normally put off, like turning a rough budget into a proper monthly plan or untangling a confusing set of instructions, and see how far it gets before you have to step in.