Kimi K3.1: The Teaser Confirms the Name, and Withholds Everything Else

Update, 24 October 2026: the model this page described as unshipped has since launched. Kimi K3.1 shipped as Kimi K3.1 Pro. The identifier appeared in Kimi Code's model switcher on 21 October, Kimi's account posted a teaser on 23 October, and the model then went live on the API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed — a 3T-class open-weight model with a 1M-token context window at $2.00 per million input tokens and $10.00 per million output, a third below Kimi K3 — with full weights promised by 4 November. The launch, and what it changes for a Kimi K3 workload. Everything below records the state of play at the teaser: what a 24-second video confirmed, and what was still missing on 23 October.

Kimi K3.1 now exists as a model name in the vendor's own voice, and as almost nothing else. The 24-second clip Kimi's official account posted on 23 October 2026 shows a retro-styled computer running Kimi Code, the lab's coding agent, with a single line above the terminal output reading "Model: Kimi K3.1" while the agent writes a Three.js scene. That is the first time Moonshot AI has put the name on a surface it controls. It is also the whole of what the vendor has said: no price per million tokens, no context window, no parameter count, no weight repository, no licence terms, and no release date appear anywhere in the clip or on any Moonshot page we checked the same day. Kimi K3, the 2.8-trillion-parameter flagship this model is a point release of, was, on that day, the newest Kimi model you could call through a public API.

That gap is worth a page rather than a paragraph, because the last three weeks have produced seven separate "signals" about Kimi K3.1 and not one of them is a rate-card row. Sorting them by what they actually prove is more useful than repeating them in date order, especially now that the strongest one has finally arrived and turns out to be a caption.

What the video shows, and what a demo cannot show

The clip is styled as an old personal computer: CRT curve, chunky bezel, monospaced type. Inside it sits Kimi Code, and the terminal header names the model. The task on screen is code generation for a 3D scene in Three.js, the browser graphics library developers have used for years as a quick visual test of a model's front-end output. Kimi's own account posted it; the outlet reports that followed on 23 October agree on the length, the retro styling, the model line and the Three.js task.

A demo like this answers one question and raises the rest. It answers "does Moonshot intend to sell a model under this name" with a yes — a vendor does not put an unpublished model identifier into the header of its own coding product's terminal otherwise. It cannot answer anything about capability. A 24-second clip is a marketing artifact, not an evaluation: it shows one task, chosen by the vendor, at a resolution that hides token counts, latency and cost. Nothing in the frame states parameters, context length, benchmark scores or a date, and the several Chinese-language reports that describe the clip as pointing at an imminent launch are inferring that from the timing rather than reading it off the video.

We could not load the original post from X to inspect it frame by frame; every detail above comes from the vendor's clip as described by several outlets that did, and they agree with each other. Treat the visual description as secondhand and the name as first-party, which is the same distinction this blog applies to benchmark numbers.

Three weeks ago the dot was a search habit. It is not one now

On 13 October, a page on this blog concluded that Kimi K3.1 was the spelling search boxes prefer for the Kimi K3 line rather than a version of it, and the evidence at the time supported that: Moonshot's documentation, pricing page, Hugging Face organisation and blog used one name for one model, and no vendor surface carried a dotted suffix. On 16 October a comparison piece went further and said K3.1 was not a model at all. Both were correct when published and are out of date now, in a specific and datable way: on 21 October the identifier began appearing in Kimi Code's own model-switching configuration, and on 23 October the vendor's account named it in a video.

What has changed is not the model's capability but its status as a noun. A search-box alias is something readers type; a name in the vendor's product is something the vendor ships. Those are different facts, and the second one is what turns a leak-watching exercise into a purchase decision — because a name in Kimi Code means there is a product line whose successor is about to be announced, even if the announcement is currently a caption.

The signal ledger, graded by what it proves

Seven traces have accumulated since late September. They are not equally strong, and two of them are routinely reported as stronger than they are.

• 28 September — a kimi-k3-1 identifier reportedly appeared in Moonshot's API backend, with an open-platform preview listing a 1M-token context window. Reported by Chinese technology press; we have not seen the page, and no public model list carries it today.

• 17–18 October — developers inspecting Moonshot's own website files found strings including k3d1-agent and make3d. First-party code on a first-party domain, which is why this one ranks high; but a front-end identifier is naming machinery, not a model.

• 19 October — Chinese outlets reported an SGLang commit titled "Kimi-K3.1 support for the K3 chat encoder," with 107 tests and a fourth reasoning level. This is the claim that gets quoted most and it is the one we could not reproduce: paging through all 538 commits in the SGLang repository between 17 and 23 October returns no commit mentioning K3.1, and a search of that repository's issues and pull requests for the string "k3.1" returns zero results. The work it describes may exist somewhere private. In public, it does not exist.

• 21 October — a developer posted a screenshot of Kimi Code's mid-session model-switch warning naming Kimi K3.1. This is the strongest pre-teaser evidence precisely because it is product-side rather than a code string: a switching UI lists models the product can route to.

• 22 October — LMArena added two anonymous entries, "Kibitz" and "Kalmia," and the community read Kibitz as Kimi K3.1. Plausible and unproven; "Kibitz" was also used as an alias in the K2 era, so the name may be a recycled one rather than a deliberate tell.

• 22 October — seven SGLang pull requests opened in a single day, five of them explicit Kimi K3 fixes for output validation, tool-call streaming and KDA kernels, the other two general MoE quantization and FP8 decode work. These are real and we verified each one on GitHub. They are also all about K3, not K3.1: none of the seven titles mentions a successor, which is the opposite of the "K3.1 is being wired up" reading they often get.

• 23 October — the teaser. First-party, dated, and the only item on the list that is the vendor speaking rather than someone reading the vendor's code.

The fourth rung: what a medium reasoning tier would actually buy

The one technical detail that has survived every retelling is a reasoning ladder. Kimi K3 sets the depth of its thinking through a top-level reasoning_effort field rather than through sampling parameters, and Kimi's own documentation lists exactly three values for it — low, high and max — with max as the default. The reported K3.1 change inserts a fourth, medium, between low and high, keeping max as the default.

If that is accurate, it is a cost lever rather than a capability jump. Reasoning tokens are billed as output, so on K3 every call that does not explicitly set an effort level is paying for the deepest setting. A middle rung would let a workload spend the top setting on the hard fifth of its traffic and something cheaper on the rest without falling all the way to low. That is the least dramatic possible change to a model and, for anyone running K3 at volume, the most useful one — which is also why it is a credible detail: it is exactly the kind of change an inference stack has to encode before a launch, because the router must render a request the model will accept.

None of this is confirmed. The vendor's documentation today lists three values and no mention of K3.1.

What is still missing, and where we looked

The check took a few minutes and the answer was uniform across every surface a buyer would consult. Moonshot's open-platform model list carries kimi-k2, kimi-k2.6, kimi-k2.7 and kimi-k3 and no successor. The pricing page's K3 and K2 tables have no row for it. The moonshotai organisation on Hugging Face has nothing newer than Kimi-K3, created on 13 June 2026 and last modified on 2 September. The Kimi Code product page still bills the agent as "Kimi Code with Kimi K3," with no mention of a newer model. And the API catalogue we route from — which mirrors the vendor's shipped models — lists four Kimi entries, none of them a K3.1.

That gives a short and unusually testable list of what a real launch looks like. A rate-card row with a price per million tokens; a model identifier that answers on the standard chat endpoint; a documentation page; and, if the K3 precedent holds, a weight repository under the vendor's organisation, since K3 shipped under open weights on 27 July 2026. The teaser covers none of those four. When one of them appears, the launch is happening; until then, coverage that describes K3.1 as "released" is describing a video.

What to do between now and the launch

For anyone already running Kimi K3, the practical answer is that nothing changes today. K3 remains the model on the rate card: 2.8 trillion parameters, a 1,048,576-token context window, image input, open weights, and a list price of $3.00 per million input tokens against $15.00 per million output. On OrcaRouter that is the price you pay — the provider's list is passed through with no markup added — and it sits on the same key as 200-plus other models, so a K3.1 row appearing later is a change on our side rather than a new contract on yours.

If the middle rung is what you are waiting for, most of the benefit is available now by setting effort per call. Extraction, classification and formatting run fine at low; analysis and multi-file edits are what high is for; max is for the debugging session you cannot afford to have fail. That routing decision is a line of request configuration, and a team that sets it deliberately will notice a larger change in its bill than a new tier will produce on its own.

For teams that want to try K3.1 the week it lands without betting a production path on an unproven model, the sensible shape is a secondary route with failover back to K3 — the new model takes the traffic you send it and the existing model absorbs the rest if it stalls. Trying a point release should not require a migration.

The date is the open question. Kimi K3 was announced in July 2026 and its weights arrived on 27 July; at the time of writing there was still one Kimi flagship on the rate card, and the teaser is the first sentence of the announcement rather than the announcement. The next thing worth watching is not another leak — it is the vendor's own pricing table gaining a fifth row.