Kimi K3.1 Pro Released: No. 4 on Global Intelligence Index, Tops Agent Capabilities

Last night, X was flooded with posts about the just-released Kimi K3.1 Pro!

Foreign developers exclaimed: welcome to a new era of 3D web development.

Some remarked: the domestic large models released recently have all been "Flash" models focused on cost-effectiveness—where are the truly capable "Pro" ones?

Many are asking: have Chinese models fallen behind?

Today, Kimi broke the silence—Kimi K3.1 Pro is here!

The well-known AI evaluation organization Artificial Analysis also published its results: K3.1 Pro shot up to 4th place on the global intelligence index with a score of 55, a jump of 11 points over its predecessor.

On Terminal-Bench 4.0, which tests real-world Agent capability, it soared from 13% to 65.7%, less than 1% behind Opus 5.5!

Artificial Analysis also made a major judgment: once the weights are released, K3.1 Pro will top the open-weight model rankings, and the gap between open-source and top closed-source models will shrink sharply from 12 points to 3 points!

This morning, Reddit was flooded with this benchmark score: Kimi K3.1 surpassed Fable 5.1!

This time, the Chinese model has truly kicked open the door to the global first tier.

The king of Chinese models returns!

What Kimi is serving up this time is not an interim version, but a serious main course.

This is the most capable 3-trillion-parameter model to date, supporting up to a 1-million-token context.

Its arrival tells the world: China is back, and the king of Chinese models has returned!

In major hardcore tests, K3.1 Pro's performance directly rivals the overseas top model GPT-6 Astra! Even more absurd, its cost is only one-fifth that of Astra—a full 80% cheaper!

So K3.1 Pro's positioning is clear: to be the most rational choice among frontier models.

Moreover, in real-world scenarios, tackling challenging tasks still requires larger models. For example, although its benchmark scores are close to Opus, in real scenarios a larger model like Fable often shows stronger insight, creativity, and writing ability.

After all, only high-difficulty reasoning tasks can show what a large model is really made of.

This time, K3.1 Pro was post-trained based on K3, and Moonshot AI significantly increased compute for reinforcement learning.

To train jointly across hundreds of downstream domains, they re-polished the entire tech stack layer by layer—from the RL framework, long-context reasoning, and sandboxes to the scheduling system and environment quality-check platform. The goal: every bit of added compute must turn into real capability.

The result: on coding, Agent, and knowledge benchmarks, K3.1 Pro reaches or approaches the level of the strongest closed-source models.

On Terminal-Bench 4.0, which tests whether an Agent can work independently in a real terminal, it scored 65.7, and at all three thinking intensities it sits on the Pareto frontier of "performance versus per-task cost."

In other words, for the same money it does the most work; for the same work it spends the least.

Kimi also thoughtfully rolled out a high-speed version, Kimi K3.1 Pro Highspeed—same intelligence, but output speed directly tripled.

But benchmark scores are only the ticket in. No matter how tightly the scores cluster on the leaderboard, the gap shows up once real work begins—

The harder the task, the more it tests the model's underlying foundation, and that is exactly where large models shine.

Next, let's see how K3.1 Pro performs in real workflows.

Coding models get a major aesthetic upgrade

Reportedly, K3.1 Pro has a striking change: aesthetics have become a capability of the model.

The question Kimi wanted to answer: can a coding model be good-looking?

This time, whether it's front-end pages, Three.js interactive scenes, or 3D assets, it shows clear improvements in composition, layout, motion, and overall polish.

It can also use game engines like Unity, Unreal Engine, and Godot, and without MCP connectors, it can write scripts itself to directly drive modeling in Blender, SolidWorks, and Rhino.

Next, let's run a round of hardcore hands-on testing to see just how capable the "big and beautiful" K3.1 Pro really is.

Hands-on: text → Blender → web page, a one-stop 3D asset pipeline

This is the flagship new skill of K3.1 Pro. Officially, it can infer shape, proportions, and spatial relationships from a passage of text, a sketch, or a reference image, directly write Blender scripts to model, and then hook into Three.js or a game engine.

Let's verify whether this pipeline really works end to end.

We threw in a hand-drawn three-view sketch, and in 40 minutes it delivered: 608 lines of Blender script, an editable .blend file, a 1.7 MB GLB, and a web page you can open with a double-click.

The sketch specified a wingspan of 8.0, total length of 6.2, and total height of 2.4; the build came out at 8.000, 6.199, 2.400—off by 1 millimeter.

Then it began its smooth operation.

Before starting, K3.1 Pro first probed Blender 5.2's new interfaces, and only wrote the formal script once it had confirmed them; the acceptance criteria were written directly as assertions—if the rotor diameter was off by even 1 millimeter, export was not allowed; after rendering, it compared the result against the sketch and revised it for one round; once the web page was done, it even opened a browser and took two screenshots of itself—only then did it count as done.

Six to-dos, checked off one by one on its own.

The fuselage is a streamlined body lofted from splines, the cockpit glass is a curved surface snapped onto the fuselage, the blades have twist, and the tilt hierarchy of the four rotors carries all the way from Blender to the browser—pull the slider to the end, and they really tilt forward 90° around the support arms.

Throughout the whole process, K3.1 Pro never touched MCP and never used external model libraries.

Halfway through writing, only 6% of the 1M context had been used. 40 minutes is not its upper limit; it's just that this sketch was only this difficult.

Even more exciting—this whole job can be done without installing Blender locally!

kimi.com and the Kimi app have simultaneously launched the 3D Make feature. Drop the same sketch and materials straight into Kimi, say "refer to these materials and build its 3D model," and it gets to work in the cloud; this time it ran for nearly another 40 minutes.

It edited scene.js, edited style.css, ran a terminal rebuild, found that two blades rotated asymmetrically, and changed them to half-integer rotation speeds; the CSS animation pauses were inconsistent, so it fixed them; when taking screenshots the scene hadn't finished loading, so it added a "ready" check before capturing.

Finally it opened the browser itself and clicked through once, reported "no script errors or missing resources," and only then handed over the version number.

What it made is right there in the window on the right—you can rotate it, change colors, and drag the tilt slider; if you're not satisfied, just change it in the conversation and it keeps revising.

With 3D Make, you can preview and edit online, publish to a web page with one click, download the GLB and take it straight to 3D printing, and even export video or Live Photos to post on social platforms.

Previously, this was the work of three people—a modeler, a front-end developer, and an operations person; now, K3.1 Pro does it all with a single input box.

And testing by many netizens also proves that K3.1 Pro's aesthetic sense is absolutely superb.

For just $4.92, it painted a portrait in a drawing application.

Someone built a palace in the clouds in Unreal Engine for less than $10, drawing lavish praise from many netizens.

The pelican-riding-a-bicycle demo—the aesthetic sense is absolutely superb.

The Wall Street full-stack analyst has arrived

But this time, K3.1 Pro directly took down the top overseas contenders.

On OfficeQA Pro v2 (complex financial data analysis based on real financial documents, testing retrieval, cross-document verification, and multi-step numerical reasoning), K3.1 Pro scored 64.4, the current state of the art (SoTA).

On Frontier Finance, which covers the entire investment workflow from target screening to profitability analysis, K3.1 Pro ranks second, a further improvement over K3.

Next, let's look at how K3.1 Pro performs in a real investment banking workflow.

It hand-built us a professional-grade financial model for Nvidia.

It doesn't simply fill in tables; it can deliver an entire Excel workbook with linked formulas end to end! It is very much like a seasoned analyst: operating drivers can be adjusted by scenario, sensitivity analysis directly shows the impact on valuation, and every assumption is backed by hard data.

You could say that K3.1 Pro is now a true "full-stack financial grunt" that understands financial logic.

The rise of Chinese models

The emergence of Kimi K3.1 Pro proves that we still hold the strength to play a trump card at any time at the big-model table.

Of course, it is not an all-around champion; in some dimensions it still lags behind the strongest closed-source models.

But what is truly exciting is that Chinese models are launching a frontal charge at the global performance ceiling of intelligence.

Cost-effectiveness determines the market; the capability ceiling determines the future.

Kimi K3.1 Pro can already go head-to-head with the world's strongest models, and even redefine the boundaries of intelligence.

And this time performance is stronger while the price is actually cheaper: one-third cheaper than K3, and only one-fifth of GPT-6 Astra.

Now, no need to wait anxiously for the ban to be lifted. Starting today, put Kimi K3.1 Pro to use.

China is back!

Editor: Solomon Aeneas

This article is from the WeChat public account "Xinzhiyuan" (ID: AI_era), author: ASI Revelation, published by 36Kr with authorization.