https://x.com/i/article/2113826834836234240
Kimi K3.1 Pro is here, a new peak in domestic model aesthetics.
Early this morning, Kimi finally released its latest model, Kimi K3.1 Pro.
As before, Kimi dropped a teaser video yesterday morning.
Finally, after more than 3 months, Kimi has brought out its latest flagship model.
Unexpectedly, while everyone had been waiting for Fable 5.5, Kimi arrived first...
K3.1 Pro's parameter scale is the same as K3's, at the 3-trillion level, specifically 2.8T, and it also supports a 1-million-token context.
Compared with K3, this one adds the Pro suffix. Judging from that naming, I think a K3.1 Flash may follow, completing the model lineup.
At the same time, K3.1 Pro is still open source, with full model weights expected to be released before November 4.
Frankly, domestic large models haven't had quite the same visibility lately.
Overseas models have been released one after another: GPT-6 Astra, 6.1 Sol, Grok 4.7, Claude Opus 5.5, Sonnet 5.5, Haiku 5.5...
So many.
This time, Kimi has finally handed in its answer, positioned as a flagship rather than a cost-effective option, to stand up against overseas models head-on.
Let's look at the basic information first.
K3.1 Pro's API pricing is 13 yuan per million input tokens, 68 yuan per million output tokens, and 1.3 yuan per million cached input tokens. The price is still flagship-level, but compared with K3 it has been cut roughly in half.
If we compare with GPT-6 Astra using the USD API prices of both, K3.1 Pro's input and output prices are each only one-fifth of Astra's. It's also a bit cheaper than Claude Opus 5.5. Cached-hit prices are the same, but regular input and output are exactly half.
In addition, Kimi has simultaneously launched a Highspeed version with the same intelligence level. The official output speed is about 2 to 3 times that of the standard version, at double the price.
For benchmarks, here is the official version.
Overall, it's essentially just behind Claude Opus 5.5 and GPT-Astra.
On Terminal-Bench 4.0, a benchmark for real terminal programming tasks, K3.1 Pro scored 65.7%, beating GPT-6 Astra's 58.2%, and only slightly behind Claude Opus 5.5's 66.4%.
On BenchCAD, which tests 3D spatial reconstruction and CAD code generation, K3.1 Pro scored 92.7%, while GPT-6 Astra scored 95.9% and Opus 5.5 scored 96.2%.
And in serious office work, the scores grew substantially. On OfficeQA Pro v2, it even achieved SOTA.
In Kimi's own hallucination-rate evaluation, the rate dropped from K3's 23.4% to 8.3%, a 65% reduction, second only to GPT-6 Astra. As everyone knows, the GPT series has always been exceptional at hallucination control, so in scientific research it often stands far ahead.
Looking at third-party AA scores, the picture is basically similar, and it even surpasses GPT-6-Astra.
It's really, really strong, truly a shining light among domestic models.
They also released several fun cases to showcase their visual and aesthetic abilities.
So I ran quite a few tests myself overnight, focusing mainly on visual directions, 3D and video motion effects.
Let me share my overall impressions first.
What surprised me most about K3.1 Pro is still its 3D capability.
In my hands-on experience, while it's still slightly behind GPT-6 Astra overall (really only slightly), in the vast majority of tasks it is already on par with GPT-6 Astra.
For video motion effects, compared with all previous domestic models, the improvement is enormous, and it can now produce some stunning videos. In this area it still trails Opus 5.5 somewhat, but given how strict anti-distillation measures have become, Kimi essentially built this purely on its own, and reaching this level is extremely impressive.
Let me show you some of the more interesting cases from our testing.
First, 3D.
For example, the 3D jelly that went viral online after GPT-6's release: an elastic, jelly-textured 3D object generated in a web page that produces real-time physical feedback.
Now K3.1 Pro can also achieve a very good result in a single pass.
Below is what GPT-6 Astra made:
For tasks heavy on physical interaction, both of these results are very good.
So I raised the difficulty.
I had it build a high-precision mechanical keyboard.
The prompt I used was:
Create a web 3D project showcasing a mechanical keyboard with extremely high fidelity. Using the same model and configuration as the baseline, faithfully and professionally reproduce the keycap surfaces, legends and grain textures, case chamfers, metal finishing, and component assembly details, and show different materials such as plastic, metal, transparent switch housings, PCB, and sound-dampening materials. It should allow free switching between bright/dark (backlit) mode, material/white model/X-ray/exploded views, and 2D orthographic top/front/left views. It should support close-up inspection, key press and rebound, hiding and isolating specified parts, and restore the internal structure as much as possible, especially the bottom of keycaps, switch housings, stems, springs and metal contacts, stabilizers and their linkages, and the assembly structure between the plate, PCB, sound-dampening layers, and case.
It should show parts such as keycaps, switches, plate, PCB, and case, and allow switching to white model, X-ray, exploded view, and different viewing directions.
Let's first look at what GPT-6 Astra produced:
Now here's K3.1 Pro's result.
You can judge for yourselves.
Comparing them, the overall accuracy is slightly behind GPT-6 Astra, but the precision of the keyboard K3.1 Pro produced has fully reached the same level as GPT-6 Astra.
Beyond 3D modeling of objects, one of GPT-6-Astra's strengths is spatial understanding, so I tested that too.
I found a floor plan and a reference image of the interior decor style I wanted:
Then I asked K3.1 Pro to use Three.js to build a website where you can roam in first person in the browser.
One image is the floor plan, the other is the renovation style rendering I want. Please use Three.js to turn it into a website that can be roamed in first person in the browser, faithfully restoring the actual proportions according to the floor plan, and making every room, every decoration detail, and every piece of furniture match the photos. The acceptance criterion is these real photos themselves. Break the house into independently assessable parts, assign an Agent to build each part, plus a reviewer Agent with a fresh context. The reviewer Agent should compare screenshots of the rooms we built with the corresponding real photos, pick the better one, point out the biggest gap, and send feedback back to the building Agent for revision. Keep repeating this process until our version wins or I stop it. Finally, tell me which parts still couldn't be restored.
This kind of task requires more comprehensive model capabilities: beyond frontend 3D modeling, it also demands visual and spatial understanding.
The floor plan only tells it how the space is divided, and the rendering shows roughly what the decor looks like. It has to build a navigable 3D scene based entirely on these two 2D images.
Below is the final result Kimi produced:
You can walk around the entire house in first person.
There's also a small minimap in the top right that shows your current position in real time.
Besides restoring the decor style and elements of the reference image quite well, there's one point that surprised me:
The decor reference image I provided doesn't correspond to any specific location in the floor plan, so the architectural skeleton inevitably differs. For example, the living room in the floor plan isn't very spacious, has a small balcony attached on one side, and is separated by walls in the middle, but the TV feature wall in the decor image is clearly much longer.
Yet K3.1 Pro can now very accurately reference the essential elements of both:
It takes the style from the decor image and the layout from the floor plan, ridiculously strong.
Finally, I tried it on our company layout.
Our company recently moved to a new office, so I also gave Kimi the office floor plan and asked it to build a model of the office.
But there was a small problem.
After moving in, we gradually added quite a few things: a sofa rest area, a dining and leisure area, plants, and so on.
So the actual arrangement already differed somewhat from this floor plan.
To make it more faithful, I casually recorded a video on my phone while chatting and gave it to Kimi.
I was curious whether K3.1 Pro could match the base map to my handheld video, even though it's the same office, with the base map giving the basic layout.
To be honest, I had very low expectations when I sent the video, but the result was much better than I expected.
The overall spatial distribution of the office, including the placement of desks and chairs and the position of the rest area, was basically all correct.
Some items that weren't in the original floor plan and were added later, such as the lazy sofa and plants, were also understood and restored from the handheld video provided afterward.
Across the entire 3D dimension, it can almost go head-to-head with GPT-6-Astra.
Beyond 3D, there's probably what everyone cares about most: K3.1 Pro's ability to create motion effects and videos.
After all, once Opus 5.5 was released, all kinds of videos made with it went viral online, and this thing is extremely demanding on both coding and aesthetics.
So with this Kimi update, I was also curious how far it could go in this area.
I first used K3.1 Pro to casually put together a motion graphics piece, using the prompt below.
Create a 16:9 landscape dynamic graphics portfolio opener, at a level that represents the work of a top motion designer. The piece should concentrate on displaying multiple advanced techniques: dynamic typography, smooth transitions, a combination of 2D and 3D animation, abstract geometric composition, fluid simulation, particle systems, distortion deformation, creative masking, multi-layer compositing, light and shadow modeling, spatial depth, and coherent camera movement. The overall rhythm should be tight, crisp, and full of confidence, with frames interlocking and transitioning naturally, constantly creating visual surprises. It should have a unified and clear art direction, with color, style, and motion language highly consistent, avoiding a mere pile-up of effects. Pursue excellence in creative conception, detail refinement, rhythm control, and visual impact, so that after watching, one immediately wants to hire this designer.
This is a first-pass result, straight out.
Objectively speaking, at this level there are still some gaps compared with Opus 5.5, but that's Opus 5.5, whose stunning quality is in a league of its own, so this is already quite impressive.
The frames are linked together with orange dot elements as transitions, so it looks fairly coherent and not like a jumble of effects flying everywhere.
Then I gave it my AIHOT website and asked it to make a simple promotional video.
To be honest, this exceeded some of my expectations: various motion effects, and even voiceover, all came out.
In terms of polish, there are some detail issues; if you really want to use it as a formal promo video, you'll need further iteration. But for a first-pass result to achieve this, I think it's already super great.
Next, there are also the explainer motion videos that went viral because of the sudden arrival of Opus 5.5.
Now you can also have K3.1 Pro make one from a single sentence.
For example, I just blurted out something very plain:
Make a hand-drawn style explainer video explaining how printers work.
And this is the video effect it produced:
If you need a more systematic and professional explanation, you can refine the request a bit.
For instance, I asked it to make an extremely detailed, professional explainer video aimed at a beginner like me, explaining information theory.
And you get an explanatory video of nearly half an hour.
Nearly half an hour, really, it's outrageous...
This is practically a whole class.
Of course, whether the explanations are accurate still needs to be checked by a human.
But being able to produce a nearly half-hour finished piece from one plain request is already remarkably impressive in terms of completeness.
Overall, after testing, K3.1 Pro has impressed me greatly in 3D capabilities and video motion effects.
Now, the Kimi web version, the app, Kimi Work, and Kimi Code can all use K3.1 Pro.
I strongly recommend everyone give it a try. This is a domestic model, and at the genuine flagship level, it's a truly fair competition.
As for Kimi's previous subscription quotas, those who know, know, and I won't complain further. It can't be helped, since domestic chips are short for everyone...
This time, alongside the K3.1 Pro launch, they also renamed all the subscription plans, now Plus, Pro, Max, Ultra and so on, basically aligned with overseas pricing, just a little cheaper.
So if you have heavy usage, it's best to buy Max or Ultra, which roughly match the overseas $100 and $200 tiers, and use the high-speed fast model sparingly.
Still, this is almost the only domestic flagship model that can be compared with Claude Opus 5.5 and GPT-6 Astra, so use it with confidence. Sometimes that's what matters most.
Finally, I'd like to share a line from Arthur C. Clarke's *Profiles of the Future* that Kimi quoted when officially announcing K3.1 Pro:
The only way of discovering the limits of the possible is to venture a little way past them into the impossible.
The only way to discover the limits of the possible is to take a small step past them, into the impossible.
When Claude is wildly banning accounts and GPT is wildly dumbing down and cutting quotas, we desperately need, and I'm very glad to see, a domestic model stepping up to arm-wrestle with top overseas models.
And at a much lower price.
To search for the outline of the future, the only way to explore the limits of the possible is to take a small step toward the impossible.
May we always stay curious.
Always have the courage to walk a little toward the impossible.
Even if it's just a small step.
Thanks, Kimi.
That's all. Since you've read this far, if you found it worthwhile, feel free to give it a like, comment, and repost, a triple combo. It helps us a great deal~
Thanks for reading my article. See you next time.