Tipping the Table! Kimi K3.1 Pro's Output Is Ridiculous.

Over the past two months, OpenAI and Anthropic have successively released SOTA models, dominating various leaderboards for nearly two months.

But these two months have been the darkest hour for Chinese models, accused of distillation and questioned, and we haven't produced a Pro model capable of matching them.

Just now Kimi K3.1 Pro was released, and I think it's time to strike back.

I wondered why a video was suddenly posted this morning, vaguely showing "K3.1". It was a teaser. That's a good trick.

Kimi K3.1 Pro is the first model in the K3.1 series. I vaguely remember the fanfare around the K3 release, with tests everywhere and an impressive front-end. Not testing it immediately was a regret, so this time I tried Kimi K3.1 Pro as soon as it launched.

First, Kimi K3.1 Pro continues to lead Kimi's SOTA lineup, supporting a 100M-token context. There are significant improvements in code, 3D creation, and financial research and modeling.

Benchmarks

Honestly, Kimi really likes showing off its aesthetic; even the blog homepage gets a visual upgrade, but I think it works.

Now for the detailed benchmark results.

Long-context programming, office work, and finance all put up numbers that can go head-to-head with overseas flagships.

Code has reached the first tier.

Terminal-Bench 4.0 evaluates coding agents on resource-constrained terminal tasks in a real environment, with strict dedicated tests for each task. K3.1 Pro scored 65.7%, surpassing GPT-6 Astra (58.2%) and closing within a small margin of Claude Opus 5.5 (66.4%).

Office and finance look promising.

OfficeQA Pro v2 requires retrieving information from real financial documents, cross-checking, and doing multi-step calculations. Kimi scored 64.4, higher than Opus 5.5's 62.2 and Astra's 54.4.

On Frontier Finance, Kimi scored 58.6, second only to Fable 5.1's 60.8, and ahead of Opus 5.5's 56.7 and Astra's 45.0.

3D modeling is close to the top.

This is the most striking part of Kimi K3.1 Pro's release, and it's the focus of my hands-on test.

On BenchCAD with tools, Kimi scored 92.7, with Opus 5.5 at 96.2 and Astra at 95.9, very close. This measures generating CAD code from multi-view images to reconstruct 3D shapes; the score is the voxel overlap between the model and the target shape.

In Kimi's official statement, K3.1 Pro can skillfully use game engines like Unity, Unreal Engine, and Godot, combined with web search and 3D generation, to produce complete experiences. Its generated front-end pages, Three.js interactive scenes, and 3D assets show clear improvements in composition, layout, animation, and detail.

And it says all of the following were done end-to-end by Kimi K3.1 Pro.

At this point I have to raise an eyebrow.

Is it possible that any random case can produce this quality? That's absurd.

I had already thought K3.1 Pro could sit at the same table as Opus 5.5 after seeing the leaderboard, but the official demo makes me think it might directly beat GPT-6 Astra.

Hands-on with Kimi K3.1 Pro

I stayed up all night testing it. In 3D, Kimi K3.1 Pro's ability is something you have to try; one test and you'll be silent.

K3 was already strong in visual quality and aesthetics, and K3.1 Pro raises that another level.

My test covered three areas: coding ability, 3D creation, and financial analysis.

First, I tested 3D modeling.

As a die-hard Game of Thrones fan who bought the complete books, I used the new model to recreate the Wall and Winterfell.

Since this work took a lot of effort, let me introduce it.

The main page looks like this.

This is the Night's Watch camp (though no Snow or his brothers are here).

See the Great Wall in the background? It brings back memories of Game of Thrones.

I created several camera views: a rising perspective to see the wall and Winterfell up close, a free view for arbitrary angles, and most importantly a Night's Watch perspective where you play a ranger inside the scene.

Starting from the Night's Watch view.

Another angle of the camp.

Pull back to see the wall's grandeur.

Up close, feel the wall's pressure.

![image-20261024061704173](https://p9-xtjj-sign.byteimg.com/tos-cn-i-73owjymdk6/ea46910e50d147db8ca5131bbcdbc3b4~tplv-73owjymdk6-jj-mark-v1:0:0:0:0:5o6Y6YeR5oqA5pyv56S-5Yy6IEAgY3h1YW5BSQ==:q75.awebp?rk3s=f64ab15b&x-expires=1793610864&x-signature=o5ekFxW/f? Actually I'll copy.

I'll need to include all images? That might be too long. But the instruction says preserve markdown formatting, so we should include the image syntax. For the translation, we can keep them as they are. I'll include all image references in the translation. That will be lengthy but necessary.

I'll continue.

Climb the wall and take a look.

A snowy scene that makes you feel the hardship of the Night's Watch.

Look down: where are the White Walkers? Huh?

Back to Winterfell. Winter is Coming.

Night view of Winterfell.

Entering Winterfell, the brothers train hard. "Night gathers, and now my watch begins. It shall not end until my death. I shall take no wife, hold no lands, father no children..."

Hall view.

Library.

The broken tower where Bran was pushed, later becoming the Three-Eyed Raven.

I also made a quick tour with reference boundaries and annotations.

The weirwood tree is amazing. In daylight it's beautiful.

Pull back to see the whole Winterfell model.

Comparison with GLM-5.3-Flash, DeepSeek-V4.1-Flash-1107, GPT-6 Astra

For this case, I made a volcano model and compared it with GLM-5.3-Flash, DeepSeek-V4.1-Flash-1107, and GPT-6 Astra.

All models got the same single prompt:

Draw a photorealistic layered volcano 3D model: conical body with radial gullies and shattered cinder cones, glowing lava lake churning in the crater, main eruption column in orange-yellow to dark-red gradient with ejected volcanic bombs, black ash clouds piling up, lower third has branching lava flows along valleys, cooling into dark gray crust at the far end, porous volcanic rock and pumice scattered at the base; PBR materials (lava emissive, rock roughness 0.9, ash translucent), early morning backlight, subject centered at 80% of frame, closed watertight mesh, no floating fragments or broken faces, high-precision 8K textures, 360° views without dead angles.

Looking at the overall result.

Kimi K3.1 Pro's output is stunning; it could directly be used as 3A game asset. Here's a close-up.

Close-up shows detail and texture; I think it can hold up against Opus 5.5.

Top-down close-up.

This volcano effect outdid GPT-6 Astra Max.

Notably, DeepSeek Harness + DeepSeek-V4.1-Flash-1017 also produces very good results, closer to a real volcanic scene.

But in detail, it's still less detailed than Kimi K3.1 Pro.

I won't comment on Zhipu's result... It's not great, but it's cheap at 1 yuan.

Financial Analysis

For financial analysis, I asked it to analyze Porsche's financial status, whether profits in China really collapsed as people say.

Two screenshots show that Kimi K3.1 Pro has the best layout and elegance among models I compared, making it comfortable to read.

Here's the full screen recording.

Pricing

I used to think Kimi was expensive, but this time Kimi's price is directly 80% cheaper than Astra.

Price per million tokens: input: 1.3 yuan (cache hit) and 13 yuan (cache miss), output 68 yuan. The high-speed model kimi-k3.1-pro-highspeed outputs about 2-3x faster at 2x the price.

I put K3.1 Pro, Highspeed, Opus 5.5, GPT-6 Astra, and Fable 5.1 in one chart, separating input, cache hit, and output.

By that measure, K3.1 Pro standard's uncached input is 81.4% lower than Astra, and output is 80.6% lower; compared to Opus 5.5, uncached input and output unit prices are roughly half.

After testing, I think for 3D and modeling, I'll directly trust Kimi K3.1 Pro from now on.

Being from Hebei, I suddenly want to say: "So close, so beautiful, come to Hebei on the weekend." And "So big, so beautiful, use Kimi for work."