Kimi K3.1 Teaser Video: What Moonshot Showed, What Is Rumor, and Where Computer Use Fits

On October 23, 2026, the official Kimi account on X posted a 24-second video that points at the next Moonshot model. In a retro-computer-style clip, a Kimi Code terminal shows the label Model: Kimi K3.1 while the model writes a Three.js 3D scene. The post passed roughly 594,000 views within hours, and replies were already calling it "Kimi K3.1 coming soon."

The video names the model. It does not date it, price it or benchmark it. This post separates what the clip shows, what earlier traces suggest, and what is still guesswork, including the idea that this is a computer-use release.

Update (October 23, 2026): Kimi K3.1 Pro is reported live, with no official benchmarks yet

Later on October 23, posts on X pointed to a model labelled Kimi K3.1 Pro rolling out in the Kimi web and mobile apps and on Kimi Code. Accounts relaying it described it as an all-around flagship for complex tasks and professional work, and one user testing it on Kimi Code reported output speeds of about 117 tokens per second. That speed is a single user's observation, not a Moonshot figure.

Moonshot has since published the official launch: see our Kimi K3.1 Pro benchmarks, price and open-weights post. The rest of this paragraph describes the state before that launch. What was still missing then: Moonshot's own announcement page, a model card, pricing, API availability and any official benchmark table. Treat "Pro" naming and the speed number as unconfirmed until Moonshot publishes them; we will add the numbers here once they appear. The earlier Kimi K3 claims are a reminder to wait for independent measurements.

TL;DR: the questions people are asking

What the video actually shows

According to the report from KuCoin News, which in turn credits MetaEra, the clip has three visible parts:

• A retro computer aesthetic, in the same teaser style Moonshot used in July.
• A Kimi Code terminal with the on-screen text "Model: Kimi K3.1."
• The model generating Three.js code for a complex 3D scene.

That is the whole evidence from the video itself. There is no spoken or written claim about parameters, context length, benchmarks or a date. We could not load the original X post directly, because X returned an error to our fetch, so the description above is secondhand from the reporting. Check the post yourself before quoting any detail.

For context, Moonshot's original Kimi K3 teaser on July 15, 2026 was also wordless. We covered how that launch played out in our Kimi K3 open weights guide and the earlier beta leaks post. The pattern is similar: a short visual teaser first, the technical detail later.

The three traces behind "K3.1"

The video is not the first sign. The reporting lists two earlier traces:

1. Late September 2026. Developers reportedly found a kimi-k3-1 identifier on the Moonshot API platform.
2. October 21, 2026. K3.1 appeared in Kimi Code's model-switching configuration.
3. October 23, 2026. The official video names the model in the product itself.

Treat the first two with care. They are reported secondhand, and the KuCoin page does not link the API record or the configuration file. A separate snapshot of Moonshot's API model list, read on October 13, listed four callable models (kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.6) and no K3.1 entry. That does not contradict the claim, since an identifier can appear and disappear, but it shows why a model name in a config is not the same as a release.

There was also a July 26 post on X labelled "Kimi K3.1 Leak" listing rumored features. It has no benchmark numbers or date, and Moonshot never confirmed it. Ignore any feature list that traces back to it.

The official video changes one thing: Moonshot itself is now showing the name in its own coding product. That moves "K3.1" from rumor to a real, named model. It does not move any spec out of rumor.

Is this a computer-use release?

The question that prompted this post was whether the video is about computer use. On the evidence, the answer is not that we can see.

Angle brackets around a dot, representing the code a model writes in a Kimi Code terminal

• The clip shows code generation in a terminal. A computer-use model would be shown operating a screen, a browser or a desktop app.
• KuCoin's page says nothing about computer use. Its only capability description is writing Three.js scenes.
• No Moonshot post, model card or documentation for K3.1 mentions it.

There are reasons the guess is reasonable. Kimi K3 is natively multimodal and reads text, images, video and documents together, per Moonshot's K3 page, which is the basic ingredient of a screen-reading agent. Moonshot also already sells a desktop agent. Our Kimi Work guide covers WebBridge, swarms and cron jobs, which is the nearest thing in the lineup to computer use today.

So the fair framing is: a computer-use upgrade is plausible, and nothing seen so far supports it. If your plans depend on it, wait for the model card. For how the open computer-use field looks right now, see our write-up of Alibaba's OpenGCUA stack.

Why a Three.js demo?

Moonshot chose a 3D scene on purpose. Three.js output is a popular informal test because the result is visual, takes many lines of coherent code, and fails in obvious ways: a broken camera, missing lights, geometry in the wrong place. It is easy to judge in 24 seconds.

It is also a marketing choice, not a benchmark. One impressive scene says nothing about the success rate over 100 prompts, and it may be the best of several attempts. Independent results are what matter. For how K3 fared on front-end tasks measured by others, see our Next.js evals and Frontend Code Arena post.

What we know about the base model, Kimi K3

K3.1 will presumably build on K3, so the K3 facts are the best anchor. From Moonshot's own page:

The page gives no price in text and shows benchmarks only as images, so verify numbers in Moonshot's tech blog. If you want the practical side of running or calling K3, our mobile app and API guide is the place to start, and the vLLM speed check shows how to read the throughput claims.

What people are asking

Should I wait for K3.1 before building on K3? No. Build on what is released. If you route through a gateway or an Anthropic-compatible endpoint, keep the model name in configuration so you can switch when K3.1 ships.

Will K3.1 break my prompts? A point release usually keeps the API shape, but behaviour can shift. Keep a small regression set of your own prompts and run it on day one.

Is there a launch date? One outlet expects a release on October 23 based on the video and "on-chain news." The page does not explain the on-chain claim, and it adds a token-listing suggestion that is pure speculation. We would not plan around it.

What about the "distilled Claude's trailers" reply? It is a joke from a user about the video's style. It is not evidence of anything. The real allegations and the political fallout are in our coverage of the White House and Moonshot distillation dispute.

Where does this sit against Kimi's earlier coding models? K2.7 Code was the coding-focused line before K3. Our K2.7 Code guide shows what that line offered, which is the baseline K3.1 would have to beat.

What this means for what you build or pay

For builders, the practical consequences are small until Moonshot publishes details, but they are worth planning for.

• Pricing. Moonshot has not announced K3.1 pricing. If it keeps K3's tier, nothing changes. If it ships as a premium model, your cost per task changes. Track the pricing page, not the rumor.
• Open weights. If weights land on Hugging Face, local and self-hosted runs become possible, as they did for K3. If not, you are API-only.
• Tooling. K3.1 already appears in Kimi Code's switcher, so Kimi Code users may see it first. Expect third-party agent harnesses to add it within days of an API release.
• Evaluation. Re-run your own tasks. A teaser that shows one 3D scene should not change a procurement decision.

A checklist for reading any model teaser

1. Is the name in a product or only in a config? A name in the vendor's own UI is stronger than one in a leaked file.
2. Does the demo have a baseline? One scene without a comparison to the previous model shows nothing.
3. What is missing? Context length, price, license, benchmarks and a date are the usual gaps.
4. Who is repeating it? Many outlets quoting one aggregator is one source, not many.
5. What would change your plan? If the answer is nothing yet, wait.

What to watch next

• An official Moonshot post, model card or tech blog entry for K3.1.
• A Hugging Face listing, and the license text on it.
• Pricing on the Kimi API page and whether kimi-k3-1 becomes callable.
• Independent evals, particularly for agentic coding and any screen-use tasks.
• Any statement on computer use, which would settle the question that started this post.

Related reading

• Kimi K3 open weights: 2.8 trillion parameters
• Kimi K3 beta leaks
• Kimi Work desktop agent: WebBridge, swarm, cron
• Kimi K3 mobile app and API guide
• Kimi K3 842 tokens/s on vLLM, claimed vs. verified
• Kimi K3 Next.js evals and Frontend Code Arena
• Alibaba OpenGCUA computer-use stack
• Sources: KuCoin News report on the video, Moonshot's Kimi K3 page

Details are accurate as of October 23, 2026 and come from secondhand reports of Moonshot's video; Moonshot had published no K3.1 model card, price or release date at that time.