Kimi K4: What Is Actually Established, and What You Can Run Today

K​imi K4 is a name with nothing behind it — that is the answer to the question most people are typing, and it is worth stating before anything else. As of 13 October 2026, M​oonshot AI has not announced, documented, priced or released a model called K​imi K4. There is no entry for it on the lab's own blog, no model card or repository in its Hugging Face organisation, no kimi-k4 identifier in the K​imi API's model list, no rate-card row and no licence. What does exist in the line today is Kimi K3, the 2.8-trillion-parameter model M​oonshot released on 16 July 2026, and it is the only model in the K series a reader can call right now.

The rest of this page is the evidence for that sentence, separated from the reporting that gets mistaken for it — because the gap between what M​oonshot has said and what third parties have inferred is almost the whole of what is knowable about K4 today.

What M​oonshot has said about K4: nothing, and that is checkable

A claim that something does not exist decays the moment it is written, so here is the method rather than only the conclusion. Every surface below was checked on 13 October 2026 and each one is repeatable in an afternoon.

• The lab's own blog and research index — the newest entry is Kimi K3, carrying the date 16 July 2026, with Kimi K2.6 below it at 20 April 2026. There is no K4 post, no teaser, and no "coming soon" line.

• The moonshotai organisation on Hugging Face — its newest text model is moonshotai/Kimi-K3, created 13 June 2026 and last modified 2 September 2026. Working down the list, the next four are Kimi-K2.7-Code, Kimi-K2.6, Kimi-K2.5 and Kimi-K2-Instruct. Nothing named K4 sits above them, and a repository search for the term returns exactly one result, evewashere/Kimi-K4, which is a third-party re-upload created 27 July 2026: 12 downloads, no likes, no documentation, and a tag set copied wholesale from K3. It is not a M​oonshot artefact and this page does not treat it as one.

• The Kimi API platform's documentation index — the model pages it enumerates are the kimi-k3 quickstart, tool-calling and agent guides, and the Kimi K2.x pages. The string "K4" does not appear anywhere in that index. The API's own model list is unambiguous: kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6.

• The vendor's pricing page — one K3 row and three K2 rows. There is no K4 row to price, which matters more than it sounds: a price is the cheapest thing a lab publishes when a model opens, and its absence dates the name more precisely than any rumour.

• Artificial Analysis — a live page for Kimi K3 headed "Released July 2026", and a 404 for the K4 equivalent.

None of that is a statement about obscurity. Kimi K3 has been public since July and leaves a checkable trail everywhere: a 2.8T model card under a custom licence, more than a million downloads on its repository, a rate card, an API identifier and an independent leaderboard entry. A shipped M​oonshot flagship leaves that trail within days. A name that has left none of it is not a quiet release. It is a name.

The single report underneath the name, and the four claims built on it

Almost every K4 page in circulation traces back to one story, and it is worth being precise about what that story said because the amplification has drifted a long way from it.

• What the lab has said: nothing. M​oonshot has neither acknowledged nor denied that a model called K4 is in development.

• What was reported: The Information, in a paywalled investigation dated around 29 July 2026, reported that M​oonshot was seeking additional Nvidia Blackwell-class GPU capacity specifically to train a next-generation model larger than K3. That is a chip-procurement story. It named no model, no parameter count and no date.

• What was added on top: trade and aggregator coverage through early August, plus Chinese-language technology outlets on 11 October 2026 reporting that a model beyond K3, put at roughly 5–6 trillion parameters, is in training and is "probably" the rumoured Kimi K4. The parameter figure in that reporting is attributed to no named source and appears in no M​oonshot document. Treat it as a number in circulation, not a specification.

• What is pure inference: a leaked celebration banner from around 27 July 2026 carrying the phrasing "push K4 to the absolute limit", which is the source of the name itself; the reading of that banner as a product roadmap; and every benchmark, context-length and pricing figure attached to K4 anywhere, all of which are extrapolations from K3 or inventions.

The one genuinely interesting thing on the timeline is a negative. Bloomberg reported on 13 October 2026 that M​oonshot has held talks to raise at a valuation of about $14 billion, with a share sale the same day. A company that raises at scale is not a company standing on an announced launch, and there is no credible reading in which a financing round in mid-October dates a model that has no artefacts in mid-October. It is context for the cadence question, not evidence about the product.

Every figure this page does not print is a figure it could not source. That is the rule the rest of the page follows.

Specifications: the honest version is that there are none

The brief for a reference page normally reads "give the specification, attributed." For Kimi K4 the true answer is shorter than usual, so here is what is unknown, stated as unknown rather than skipped.

• Class — unknown. M​oonshot has published no architecture, no layer count, no routing scheme and no statement of what the model would be for. Any claim that K4 is a mixture-of-experts model is inherited from K3, not sourced to K4.

• Size — unknown, and actively contested. The only figure in circulation is the 5–6 trillion parameters reported by Chinese outlets on 11 October 2026, unattributed to M​oonshot and unreproduced elsewhere. This page states it because readers will encounter it, not because it is established.

• Context window — unknown. The 1,048,576-token figure belongs to K3 and to nothing else.

• Modality — unknown. K3 accepts text and image input and returns text; whether a successor matches or widens that is not something anyone has published.

• Licence and weights — unknown, and this is the one to watch most closely. K3's weights are published under the Kimi K3 License, a custom "other" licence rather than a standard open-source one. Whether a successor inherits those terms, narrows them, or does not open at all is the question that decides more for a builder than any benchmark.

• Price and release date — unknown, with no announced date of any kind. A prediction market on K4 shipping in 2026 has been pricing a first-quarter 2027 outcome above a 2026 one since the summer, which is a measure of sentiment among bettors, not a source about M​oonshot.

Kimi K4 is also not a variant, not a quantisation and not a fine-tune: there is nothing for such a thing to be a variant of.

K3, K3.1 and K4: three names doing three different jobs

These three get blurred together in comments, and blurring them is what makes the leak reporting look more solid than it is. They are three separate objects. Kimi K3 is a shipped model with a card, weights and a price. Kimi K3.1 is a version number that has appeared only in leaked attribution, decoding games and third-party reporting about a mid-October window — no card, no weights, no identifier, and no M​oonshot confirmation on the day this page was written. Kimi K4 is a forward-looking name attached to a chip-procurement report and a banner.

There is a specific trap in treating the first two as one story. Because the spelling kimi k3 carries more search demand than the vendor's own kimi k3.1 label, and because a K3.1 report was circulating before K4 ever was, "K3.1" gets read as a stepping stone to K4. Nothing published supports that ordering. One is a numbering question inside the K3 line; the other is a next-generation question with no numbering at all.

Our own coverage sits in the same three boxes and the two sibling pages are the ones to read next: what the K3.1 attribution actually is, and what the K4 leak claims. This page is deliberately the third object and not a fourth leak page — it exists to say what is established, which is why it is short.

What would settle this

Watch for artefacts, in this order, because each is cheaper for the lab to produce than the one after it.

• A model identifier. A new string in the Kimi API's model list is the earliest hard confirmation and needs no announcement. If a kimi-k4 id shows up in /v1/models on any platform, the question of whether K4 exists is answered.

• A rate-card row. Pricing appearing next to the name means the model is being served, quoted and billed, whatever the blog has or has not said.

• A model card and repository under moonshotai on Hugging Face. This is the one that settles the numbers — parameter counts, layer structure, context length, modality and licence terms — and it is the only source from which any of them should be quoted. For this specific model the field to read first is the licence tag.

• A weight release. For anyone who needs to run the model on their own hardware rather than buy tokens, this is the step that changes what they can do, and it usually follows the card by days.

Until at least the first of those appears, the accurate answer to "is Kimi K4 out" is no, and the accurate answer to "when" is that nobody outside M​oonshot knows. This page does not carry a release date, a "coming soon", or an expected window, because there is no source for one and a wrong date is worse than an absent one.

What to run in the meantime

For a reader whose actual problem is a long-context job this month, the decision does not involve K4. Kimi K3 is the model that exists: a 2.8-trillion-parameter mixture-of-experts design that activates 104B parameters per token across 93 layers, with 16 of 896 experts firing per token, a 1,048,576-token context window and native image input, published under a custom licence with weights available. M​oonshot prices it at $3.00 per million input tokens, $0.30 cached and $15.00 per million output tokens, and it exposes a top-level reasoning_effort control rather than sampling parameters.

On OrcaRouter, kimi/kimi-k3 is live with the full 1M-token context and the provider's list price passed through with no markup, so a vendor price change on it reaches your bill the same day rather than at a repricing cycle. The same endpoint carries the rest of the line — kimi/kimi-k2.7-code and kimi/kimi-k2.6 — so comparing a coding-oriented Kimi against the flagship is a model-string change rather than a second integration. When a K4 identifier does eventually appear, the migration is a string in a routing policy: one O​penAI-compatible endpoint over 200+ models, automatic failover when a provider degrades, and a routing DSL that lets you send a slice of real traffic to an unproven model with the proven one as fallback.

To be explicit about what that is not: no route claim is made for Kimi K4 on this page. We hold no model id matching kimi-k4 or kimik4 under any spelling in our catalogue, verified against the live model list on 13 October 2026, and this page makes no claim about when one might appear. Our Kimi ids today are kimi-k3, kimi-k2.7-code, kimi-k2.6 and kimi-k2.5. That is the whole list.

The short version

If you came here to find out whether Kimi K4 exists: it does not, in any form a reader can point at. No announcement, no model card, no weights, no identifier, no price, no date — and a single paywalled chip-procurement report plus a leaked banner is the entire public basis for the name. If you came here to find out what to do about it: nothing, this month. Kimi K3 is the model in this line that runs, prices and downloads today, and the thing to watch is not another post about K4 — it is an identifier, a rate card, or a repository with the name on it.