Kimi K3.1 suspected to appear on Arena, official video reveals model name

This morning, Kimi's official account suddenly released a 24-second video.

At the start of the video, an old-fashioned computer screen is running Kimi Code, and in the terminal a piece of code is being written that uses Three.js to generate a complex 3D scene.

Observant netizens quickly noticed that the model field in the upper-left corner of the screen clearly read Kimi K3.1.

Just in the early hours of October 22, a mysterious model named Kibitz appeared on Arena, and netizens suspected it was K3.1.

Some netizens began comparing it with top models such as GPT-6 Astra Max and Claude Opus 5. One of the tests required the model to use only a single HTML file to create a 3D voxel courtyard containing a pagoda, trees, and cherry blossoms.

Kibitz delivered a colorful 3D courtyard, and the result was quite good; in the same test, although GPT-6 Astra Max also completed the work, it took about three times as long as Kibitz.

The left image is GPT-6 Astra Max, and the right image is Kibitz.

The AI model tracking account @aitrackerbot on X found through a Tokenizer test that Kibitz has obvious characteristics of Moonshot's Kimi model family. Meanwhile, another anonymous model, Kalmia, which appeared on Arena at almost the same time, is suspected to come from Alibaba's Qwen team.

Kimi K3.1 seems to really be coming soon.

How does Kibitz perform?

In Arena's Code mode, many netizens have already grabbed Kibitz to run all kinds of bizarre tests.

One classic item is having the model use SVG to draw a pelican riding a bicycle (this was originally a prompt often used by the developer community to test a model's drawing and coding abilities; almost every new model has to draw one first).

Kibitz directly turned it into an interactive webpage. In the image, the pelican is pedaling, and users can adjust the pedaling frequency and even switch between different scene effects such as daytime and dusk.

Other netizens had it create a 3D helicopter, requiring it to display the fuselage structure and switch between exploded-part views and wireframe mode. Kibitz completed an interactive 3D page in less than 15 minutes.

When the task was changed to creating a medieval-style castle, Kibitz likewise delivered a rather beautiful piece of work.

However, the community's evaluations were not uniformly positive.

Some netizens ran a side-by-side test comparing Kibitz, Kalmia, and Gemini 4 Carbon with Claude Opus 5.5. As a result, among these models, the tester favored Carbon.

In the 3D courtyard test at the beginning, although Kibitz generated faster, the tester said that GPT-6 Astra Max's work was equally beautiful and also achieved effects such as day-night alternation and changes in light and shadow.

These scattered works are not really rigorous evaluations of model capabilities, but for a model that has not yet been officially released and whose true identity is not even certain, its performance is already enough to attract developers' attention.

As for why netizens suspect it is K3.1, there is another set of clues.

Previously, Kimi K2.5, K2.6, and K3 were tested under anonymous codenames such as Kiwido, Kiwire, and Kivine respectively, and the name Kibitz happens to continue this pattern.

Moreover, the model tracking account @aitrackerbot mentioned at the beginning also found 11 special Token strings with characteristics of the Moonshot model family.

When the model was asked to repeat these strings verbatim, Kibitz made errors in two consecutive rounds, performing similarly to other tested Kimi models (by contrast, control models from seven other labs were all able to repeat them correctly). This means that Kibitz is very likely technically connected to the Kimi model family.

In addition to these community tests, Moonshot's own product interfaces have also revealed some clues.

Just one day before Kibitz appeared on Arena (October 21), netizens discovered that the name k3.1 appeared directly in Kimi Code's configuration.

However, that evening, when netizens requested this configuration again, they found that it could no longer be obtained, and the response was directly a 404.

K3.1 should be coming too

Speaking of which, more than three months have passed since the previous generation, Kimi K3, was released.

In mid-July this year, Moonshot officially launched Kimi K3. This model has 2.8 trillion parameters and a 1 million Token context window, making it the open-weight model with the largest parameter scale in the world at that time. After its release, K3 once topped Arena's front-end programming leaderboard, even surpassing Claude Fable 5 and GPT-5.6 Sol.

Now, three months later, K3's popularity has gradually settled, and it is time for Moonshot to launch a new version.

In fact, as early as September, news about Kimi K3.1 had already begun to circulate.

At the end of September, developers discovered that the identifier kimi-k3-1 appeared to show up in Moonshot's API model registration information. Subsequently, reports said that a new model codenamed K311111 briefly appeared on the Kimi open platform, supporting a 1 million Token context.

In mid-October, netizens dug up more from the front-end code of Kimi's official website: among them were identifiers such as k3d1-agent, make3d, AUTOMATION_K3, and OK_COMPUTER_SWARM.

These names look quite interesting, especially make3d, which easily brings to mind that Kimi may be preparing 3D creation features; while AUTOMATION_K3 and OK_COMPUTER_SWARM point to automated tasks and multi-Agent collaboration.

Around October 20, another developer discovered that a code commit titled Kimi-K3.1 support for the K3 chat encoder appeared in the SGLang GitHub repository. This code directly contains moonshotai/Kimi-K3.1, and also includes compatibility tests for the model's dialogue format, tool calls, and Token encoding.

According to the commit description, K3.1 continues to use K3's Tokenizer and 16 special Tokens, and one fairly clear change is the addition of a medium reasoning intensity—previously K3 supported three reasoning levels, Low, High, and Max, while K3.1 may add a fourth level, allowing users to control the model's thinking intensity more flexibly.

However, this is mainly adaptation information for the inference framework, and is not enough to show how much the new model's overall capabilities have improved.

Judging from K3's previous positioning, Coding and Agent are likely still the focus of K3.1. If Kibitz really is K3.1, then its performance on Arena at least shows that the new model has already delivered good results in front-end and 3D scene code generation.

Another direction worth watching is inference efficiency and cost.

Previously, although K3's pricing was not low, at $3 per million input Tokens and $15 per million output Tokens, it was still about 70% cheaper than Claude Fable 5's $10 and $50. Considering its performance on Coding and Agent tasks, for developers it already counted as a fairly cost-effective choice.

But now, overseas model vendors have pushed API prices to a rather exaggerated level. Claude Opus 5.5's input and output prices have already dropped to $4 and $20 per million Tokens, and Sonnet 5.5 is only $2 and $10.

Under these circumstances, if K3.1 cannot keep actual usage costs within a range where developers are willing to keep paying, then no matter how impressive its benchmark scores are, its appeal will be greatly reduced.

In the latest official teaser video, K3.1 directly demonstrated a task of building a complex 3D scene using Three.js, echoing the previously exposed make3d feature.

Netizens therefore speculate that 3D creation is very likely one of the key upgrades in K3.1 this time. How many more new things there are will have to wait for the official release.

First, let's play a round of guess who I am

In the large model circle, anonymous leaderboard testing has almost become an "old tradition."

Vendors put a new model in a disguise and quietly place it on platforms such as Arena and OpenRouter, while netizens busily run benchmarks, analyze Tokenizers, and study output habits, trying to uncover the real identity behind it.

OpenAI already played this game in 2024.

That year, a mysterious model called im-also-a-good-gpt2-chatbot suddenly appeared on Arena, performing so well that it once shot to first place on the leaderboard. Netizens guessed for a long time, and in the end OpenAI announced: it was actually GPT-4o.

In 2025, a mysterious image model named Nano Banana suddenly went viral, and netizens tested its astonishing image editing capabilities while guessing which company had made it.

In the end, Google came forward to claim the banana, whose official identity was Gemini 2.5 Flash Image.

Google later revealed that the name Nano Banana had been improvised by a product manager at 2:30 in the morning. Unexpectedly, the model became popular, and the banana also became an official product brand.

Domestic vendors have also already mastered this game.

On August 20, a mysterious model called Ox Alpha (dubbed the "Niu Lai large model") appeared on OpenRouter and OpenCode. Not only was it free, but its Coding and Agent capabilities were also quite good, and it quickly attracted a large number of developers.

Six days later, Zhipu officially revealed the answer: it was GLM-5.3-Flash.

According to data later disclosed by Zhipu, Niu Lai's anonymous test over six days cumulatively processed more than 62 trillion Tokens, and at one point became the most-used model on the two platforms.

On September 23, another anonymous model codenamed Space Bunny landed on OpenRouter, also open for free testing and supporting a 1 million Token context.

It was not until October 19 that MiniMax finally came forward to claim it, announcing that it was an early version of M3.1 Flash.

For model vendors, anonymous testing is indeed quite worthwhile—it allows them to collect feedback from real users in advance, see whether the model can withstand large volumes of calls, and also lets developers actively help test, share their work, and guess identities.

Before the model was even officially released, the internet had already been buzzing for several rounds, and by the time the vendor came forward to claim it, most of the publicity work was already done.

However, this playbook occasionally produces a bit of awkwardness.

Some models perform stunningly while anonymous, and netizens rush to guess which lab has pulled off another big move. But once the official release arrives and everyone can use it, some people feel it isn't quite as impressive as before.

Just after guessing whose model it was, netizens began to wonder: is this really the same model as before?

Although this is understandable—what spreads widely online tends to be the most polished work—anonymous testing raises everyone's expectations too high in advance, so when users actually get their hands on it, a gap is inevitable.

Now, Kibitz and Kalmia have also joined this "game."

As for whether the alt-account guesses are correct, and whether it will still be this stunning after the official release, I think we'll find out very soon.

This article comes from the WeChat public account "Letter AI," author: Yuan Xinyue, published by 36Kr with authorization.