Kimi K3.1 Pro Released, Ranks Fourth Globally on Artificial Analysis Intelligence Index

By | Digital Force Field, Author | She Zongming

Against the backdrop of domestic large models having gone quiet for over two months, giving rise to widespread doubts of "are Chinese models already finished?", Kimi stepped up with a move.

In the early hours of October 24, K3.1 Pro officially took the stage.

If K3 represented the "ceiling of open-source models" in the previous version, then K3.1 Pro is the evolved K3 that advances N more steps even at the hundred-foot pole.

Whether it can fight, its performance speaks for itself.

Artificial Analysis ranked its intelligence index fourth globally, and also offered a judgment: once the weights are released, K3.1 Pro will top the open-weight model leaderboard, and the gap between top open-source models and closed-source models will shrink from 12 points to 3 points.

▲On the Artificial Analysis leaderboard, K3.1 Pro's intelligence index ranks 4th among global large models.

On Terminal-Bench4.0, the industry-recognized hardcore terminal Agent coding benchmark, K3.1 Pro scored 65.7, second only to Claude Opus5.5's 66.4, further narrowing the gap between open-source and closed-source models.

Across the ocean, it was exceptionally lively: in September, OpenAI successively released GPT-6 Astra and GPT-6.1 Sol, Anthropic brought out Claude Opus 5.5, and the next month, even Google, previously considered to have fallen behind, returned to the table with Gemini 4 Argon...

English text is being prepared.

On the domestic model side, it formed a stark contrast: since Kimi K3 set off a global tech circle explosion, domestic models have been somewhat quiet. Model vendors did make moves to launch new models, but the volume they stirred up was almost negligible.

In short: people hoped for the bull to come, but the bull did not come.

▲Comparison of China-US large model releases over the past three months. The scores in the figure are Artificial Analysis's intelligence index ratings for each model.

As a result, the Top10 large models on multiple authoritative leaderboards have nearly become an internal roll call of American AI models.

No fear of silence, only fear of comparison. Along with the strong contrast came the rise of "distillation accusations" and "capability doubts".

"Distillation" has always been a dart that Anthropic and others throw at domestic models from a moral high ground, and especially so in these two months. From the three major agencies NSA+CISA+FBI taking action, to Anthropic yet again filing a complaint, to OpenAI publicly releasing an anti-distillation action report for the first time, all came with highly targeted aim.

And the domestic models' silence occurred after Anthropic and others strengthened their anti-distillation defense mechanisms, making it easy to be seen as "confirming it in disguise".

After Kimi K3 greatly closed the technical generation gap between top open-source models and top closed-source models, shrinking the "time gap" from over 1 year to about 4 months, the contrast of these two months has made many people doubt the judgment that "open-source models can catch up with and surpass closed-source models", feeling that the gap has been widened again.

Some even concluded that open-source model breakthroughs rely on distillation, and that anti-distillation is tantamount to pulling away the ladder that open-source models climb up.

I still remember that when K3 first came out, I wrote an article titled "Taking leave of Silicon Valley without a backward glance, half for DeepSeek and half for Kimi". But I must say, the situation over the past 3 months gave me the worry of "I won't get slapped in the face, will I?"

Against this backdrop, K3.1 Pro's "rescue" came at just the right time. It invisibly sent a signal: past the gray moment, China is back.

02

The default logic of the "distillation theory" is: distillation can help you accelerate catching up, but it cannot help you become SOTA—distillation is a shortcut for following, not a path for leading.

A relevant person in charge at Moonshot AI previously responded: the core of K3's performance leap comes from underlying original architecture innovation, not distillation replication of existing models. And those key innovation points on which K3's breakthrough depends—Moon Clip second-order optimizer + KDA hybrid linear attention + Attention Residuals + Stable LatentMoE sparse expert framework—are clearly not distilled.

This time, K3.1 Pro still achieved an upgrade under the premise that Anthropic and others launched their "anti-distillation grand method", providing the most intuitive counterexample to the claim that "domestic models rely entirely on distillation".

Under the constraints of not having an advantage in compute power and funding, whether the gap between domestic models and Silicon Valley frontier models will widen or narrow next is also a question many people care about.

It should be noted that since late July, the models released by domestic model vendors have all been Flash versions (high-speed, inclusive versions), with no Pro versions (flagship versions). Compared with the Big Three's successive flagship releases, their momentum is somewhat weaker. K 3.1 Pro fills the "Pro version gap" of domestic models over these months.

This is not to say there is a good-or-bad distinction between Pro and Flash versions, but from the performance dimension, the two do differ in strength: generally speaking, Pro corresponds to the "capability ceiling first" route; Flash corresponds to the "efficiency/cost first" route.

While other vendors have been rolling out Flash versions one after another, Kimi is still using the Pro version to probe the upper limit—which inevitably makes one want to cue Ms. Liang Jingru again.

▲Work produced by a professional overseas player using K3.1 Pro.

K3.1 Pro is equivalent to continuing the momentum of open-source models "seeing the best and striving to match it." Considering that the compute resource constraints are vastly different from those of Silicon Valley model vendors, narrowing the intelligence index gap between top Chinese and American models to 3 points is truly not easy.

What's more, K3.1 Pro is also very much on point when it comes to aesthetics, showing Taste as consistently as ever.

03

From K3 to K3.1 Pro, Kimi has proven that its technological iteration is sustainable.

From the perspective of technological innovation, achieving one stunning breakthrough and establishing a stable breakthrough mechanism are two different things.

The former may be the product of a flash of inspiration plus a bit of luck; the latter means the company has found a complete chain from inspiration, experimentation, and engineering to product—a chain that can withstand the test of time.

If Kimi had run out of steam after launching K3, then K3 would be a lone peak, but the emergence of the "sequential version number" K3.1 Pro shows that what K3 connects to is a mountain range.

In fact, the continuity of Kimi's innovation and iteration has long been traceable—

When K2 came out, it was praised by Jensen Huang, commented on by Musk, and regarded by the Nature website as "another DeepSeek moment";

After K2.5 came out, it was "paid tribute to" by Cursor, whose founder publicly called K2.5 "the strongest base model" in his evaluation, and Musk also stamped it with "Yeah, it's Kimi 2.5";

The subsequent K2.6 was the first to implement a large-scale parallel Agent cluster + browser-resident background Worker, and was also highly recognized in programmer communities...

This continuity is very important for model vendors. Many model vendors today have a pulse-like innovation curve: they save up for a big move, dominate the headlines for a while, and then... there is no then.

This inevitably brings to mind the plot of the movie "Little Monsters of Langlang Mountain": the little pig demon obtains a big move passed down from his father, but unfortunately he can only use it once in his lifetime.

▲Still from the movie "Little Monsters of Langlang Mountain."

Looking at the long term, AGI is a technological long march measured in decades.

Runners who love marathons know that in a marathon you cannot sprint too hard in the first half and then lose steam in the second half; those who truly finish are the ones with steady pacing and lasting endurance.

The AGI journey is the same. Since it is a long run, long-term resilience matters more than brief sharpness.

04

When it comes to Moonshot AI, many people recall that these four characters are a tribute to the album of the same name by the rock band Pink Floyd.

On reflection, "the dark side of the moon" inherently carries the meaning of "half light, half unknown." Applying this meaning to Kimi is also quite apt: its path here has been full of light, while its path ahead is full of the unknown.

Looking at it now, the marathon is far from over. But it is clear that Kimi is among the front-runners.

Bruce Lee once wrote in his notes: Jeet Kune Do makes us, once we have set our direction, never look back.

Once the direction is clear, keeping moving forward is very important.