Kimi K3 and K3.1 Pro for web design

Kimi K3.1 Pro is Moonshot AI's newest model, released on October 23, 2026 as a post-trained version of Kimi K3, with a 1M token context and open weights. On Design Arena's Website board at the time of writing, Kimi K3 is fifth at 1332 and K3.1 Pro eighth at 1320, a few points from Claude Opus 5.5 and GPT-6 Astra.

That places Kimi among the models that design best by public vote. The distance to the top is a few votes in a hundred: the Website board's leader is expected to beat Kimi K3 in about 55 of 100 head-to-head votes. Moonshot's own launch post makes its strongest claims for 3D and games, and the boards agree: K3.1 Pro is fourth on Design Arena's 3D Design board, winning 71% of its 362 battles.

The short call:

• Building a one-page site in a single file? K3 and K3.1 Pro are both in the top ten; K3 edges it by 12 points, which is inside the noise.
• Building 3D scenes, games or multi-file apps? K3.1 Pro, which leads K3 on all three boards.
• Iterating many times in an afternoon? Note the time: K3.1 Pro averages almost 20 minutes per site on the Website board.
• Want the page to stop looking generated? Hand it a full specification. Thanor writes those, one per design.

What Kimi K3 and K3.1 Pro are#

Both come from Moonshot AI and share one base. The facts below are from Moonshot's K3 post, its K3.1 Pro post, its API platform and the Hugging Face model card.

K3.1 Pro is the first model of a K3.1 series, and Moonshot calls it its most capable 3T-class open-weight model. The Highspeed variant carries the same model at two to three times the speed for $4.00 and $20.00 per million tokens. Prices are as listed at the time of writing.

Where you can run it matters for a design workflow. K3.1 Pro is on kimi.com and the Kimi apps, in Kimi Work on desktop, in Kimi Code (picked with the /model command) and through the Kimi API, where Moonshot offers zero data retention to API customers. Enterprise admins have to switch it on, since it is off by default there.

The open weights are the part no closed flagship offers. A team can host the model itself, though the license is Moonshot's own rather than MIT or Apache, so read it before a client build depends on it.

Where Kimi ranks for design#

The public boards are where the claims meet votes. Design Arena figures were read on November 9, 2026; Arena's WebDev figures carry its November 8 date.

Two things stand out. K3 was the story in July, when Arena put it first on its Frontend board at launch with a 76% pairwise win rate, against 63% for Claude Fable 5 and 58% for GPT-5.6 Sol. Since then newer models from Meta, OpenAI, Anthropic and Thinking Machines have moved past it on most boards, and K3 now sits in a cluster rather than at its head.

The second is where K3.1 Pro pulled away from K3. On the 3D Design board it gained 71 points, the largest move between the two on any board above, which matches where Moonshot put its claims. A gap of 71 points means K3.1 Pro is expected to win about 60 of 100 votes against K3 there, while on the Website board the two are within a coin toss of each other.

The cost of all that is time. On the Website board K3.1 Pro averages 1,174 seconds per page and K3 836, against 309 for Claude Opus 5.5. When 3D is the point of the page, these three show what a full specification of the technique looks like:

What Moonshot says Kimi does well#

Moonshot's claims cluster around visuals and 3D. For K3, its post says the model uses screenshots and visuals to improve game development, frontend work and CAD, iterating between code and live screenshots of what it built. For K3.1 Pro the post goes further: it "brings a more considered visual style to games and web experiences", with front-end pages, interactive Three.js scenes and 3D assets that show stronger composition, typography, animation and polish.

The supporting numbers are Moonshot's own and mostly about 3D:

• BenchCAD, 3D reconstruction through CAD code: K3.1 Pro 92.7 with tools, against 96.2 for Claude Opus 5.5 in Moonshot's own chart.
• Parametric modeling, which Moonshot singles out: mechanical parts, architectural models, configurable products and furniture.
• 3D Make on kimi.com, where users preview and edit generated 3D models.

The K3 post is also unusually frank about limits, and three of them bear on web design. It says K3's overall user experience trails Claude Fable 5 and GPT-5.6 Sol. It warns that generation quality can become unstable if a harness does not pass back the model's full thinking history, or if a session switches to K3 halfway. And it says that when intent is ambiguous, K3 may make decisions on the user's behalf that the user did not expect.

That last caveat is the whole case for a full brief. A vague design prompt is ambiguous intent by definition, and an eager model fills it with choices of its own. The K3.1 Pro post does not repeat the list, so treat the caveats as K3's and check whether they still apply in your own runs.

How to brief Kimi for a designed page#

A brief for Kimi carries the same decisions as a brief for any flagship, the ones a visitor sees first, plus one more that matters for Kimi in particular: the technique.

Name the 3D technique#

Moonshot's showcase includes interactive Three.js scenes, so a prompt that only says "a 3D shelf" leaves the model to pick between a WebGL scene and something lighter. Octavo's brief asks for CSS 3D transforms moved by one script, with no WebGL at all; Kickflip's asks for hand-written WebGL on one canvas with no 3D library. Both are 3D. They weigh, load and fail differently, and the brief is where that is decided.

Write values, not adjectives#

A fragment of the kind that changes the output, written for this article:

build     one HTML file, one style block, one script, no framework
3D        CSS 3D transforms only, no WebGL, no Three.js
palette   ground #F4E9D8, ink #241A12, accent #E8432B on the one control that opens
type      display 104px at 0.9, body 17px at 1.6, ratio 1.25, two weights
motion    rest cubic-bezier(0.2, 0.8, 0.2, 1) over 600ms; hover lift 6px in 180ms
refuse    no gray neutral, no gradient, no shadow except under the object in front

Keep the history and the brief together#

Moonshot's own advice is to pass back the full thinking history. If you drive the API from your own harness, keep it, and paste the token block again at the end of a long request so the last section obeys the first. The Gemini 3 write-up explains why restating the block costs sixty words and saves the footer.

Where to watch Kimi's output#

No model is free of defaults, and Kimi's documented behavior points to three places to look when a page comes back.

1. Decisions you did not make. Given room, the model chooses. Check the typeface, the neutral and the accent against the brief before anything else, because those are the slots every model fills from the same median, as the pillar on model choice sets out.
2. The technique it reached for. If you asked for a hero and got a full Three.js scene, the page now carries a renderer you may not want. Name the technique or accept the model's.
3. Time per pass. At almost 20 minutes per site on the Website board, K3.1 Pro rewards a brief that is right the first time. The Highspeed variant exists for iteration.

Motion is where a written specification shows most clearly, because a mechanism can be described as physics. These two tie their entire page to one input, the scroll in one case and the pointer in the other:

Neither is an effect a model adds on its own. Each one is a sentence of mechanism and a set of numbers, which is the form of instruction a model can follow without guessing.

Kimi next to the rest of its tier#

Kimi's real competition for a design job is the group it shares the Website board with. Same board, same day, same voters:

On a single page the five are within 27 points, which is close to a coin toss between neighbors. On 3D the spread opens to 119 points, and only Claude Opus 5.5 sits above K3.1 Pro among them. Kimi's case is 3D at an open-weight price, and its cost is the clock.

The price side, as listed by each vendor at the time of writing: K3.1 Pro at $2 in and $10 out per million tokens, Claude Opus 5.5 at $4 and $20, and Gemini 4 Argon at an introductory $2 and $10 until November 30. The DeepSeek and Gemini models have write-ups of their own: DeepSeek V4 and Gemini 4 Argon. For what Opus 5.5 does with a long build, read the Opus 5.5 write-up.

None of those differences decides what fills a blank slot. Given only "a modern landing page", the open choices get filled from the shared median, the neutral sans and the gray near the axis that make generated sites look alike, whichever model fills them. The table tells you which engine to rent; the brief tells it what to build.

The verdict on Kimi K3.1 Pro#

Kimi K3.1 Pro is a credible top-ten design model, open-weight, and cheaper per token than Claude Opus 5.5 at the time of writing. It is the strongest Kimi for 3D, games and agent-built apps, while on a one-page site it trades places with K3 inside the noise.

Ranked by what changes the result:

• The brief: values, technique, media and refusals decide whether the page looks directed.
• The technique you name: CSS 3D, WebGL or a film, each a different page.
• K3.1 Pro over K3: a clear gain on 3D and apps, a tie on single pages.
• Kimi over another flagship: price, open weights, 3D strength and where you work, against a slower clock.

For how Kimi stacks against the rest of the field on the same boards, read Design Arena, explained. For the closest open competitor, read DeepSeek V4 for web design. For 3D pages to aim at, the best 3D websites and Three.js website examples show what a named technique produces.

Give Kimi a finished brief#

Kimi rewards a brief that is complete on the first pass, because every pass costs minutes. The fastest way to see that is to run one.

Open Vesperion and copy its prompt; it is free with an account. Paste it into Kimi Code with /model kimi-k3.1-pro, or into kimi.com, and compare the page that comes back with the moving preview. The prompt names the palette, the type, the motion and the media, so whatever differs is the model's doing, and that is the comparison worth making.

Then browse the library for a design in the category you build, the 3D shelves and turntables included, and give Kimi the decisions instead of the adjectives. If you also work in Claude, ChatGPT or Cursor, the Thanor MCP pulls the same art direction into those sessions without the clipboard.