Kimi K4

What is actually known about Kimi K4?

Nothing. Moonshot AI has not announced a model called Kimi K4, has not published specifications or pricing for one, and has not confirmed that it exists. Every field in the table above is empty by design rather than by omission — we would rather show you a dash than a number we cannot trace to Moonshot.

That is unusual for a page like this, and it is deliberate. Most pages you will find about unreleased models fill the gap with leaks, extrapolated benchmarks and invented pricing tables. We do not, because a fabricated specification is indistinguishable from a real one until it is proven wrong, by which point someone has planned around it.

What we can offer instead is the observable record: what Moonshot has actually shipped, how quickly, and what the current flagship already does. Everything below is either a verifiable fact or clearly marked as our own inference.

What has Moonshot AI actually shipped, and how fast?

Five model lines in roughly a year, and the gaps between them have been getting shorter. These are the release dates Moonshot has published, taken from the same registry that drives every other page on this site.

Model	Released	Gap since previous
Kimi K2	July 11, 2025	—
Kimi K2.5	January 2026	~5.7 months
Kimi K2.6	April 20, 2026	~3.6 months
Kimi K2.7 Code	Not published	—
Kimi K3	July 16, 2026	~2.9 months

Two honest caveats on this table. Moonshot published the month but not the day for Kimi K2.5, so its interval is approximate. And Moonshot has never published a release date for Kimi K2.7 Code at all, which means there is at least one release in this sequence the table cannot place.

Our inference, clearly labelled as such: the interval between confirmed releases has narrowed from roughly six months to roughly three. If that continued, a successor to K3 would be due in late 2026. That is an extrapolation from three data points, not a schedule, and Moonshot has committed to nothing. Treat it as a reason to keep your model identifier in configuration, not as a date to plan around.

Will the next Kimi model even be called K4?

Quite possibly not — and since this site is named after it, we would rather say so plainly. Moonshot’s naming history is K2, K2.5, K2.6, K2.7, K3. Point releases outnumber major versions four to one.

On that record, a Kimi K3.5 or K3.6 is at least as likely as a K4. The jump from the K2 series to K3 came with a substantial change — a step up to 2.8T parameters and a 1,048,576-token context window — which suggests major numbers are reserved for architectural shifts rather than issued on a schedule.

We named the site for the model people search for, and we cover the whole family regardless of what arrives next. If the successor is called K3.5, it gets the same treatment on the same day, and this page will say so.

What would a successor actually need to improve?

This section is our analysis, not reporting. Nothing here is a leak or a claim about Moonshot’s plans — it is a reading of the gaps visible in the current lineup, which is something you can verify against our own model pages.

Four things stand out as unfinished business:

• Video input on the flagship. K3 does not document video input, while Kimi K2.6 and Kimi K2.7 Code both do. A flagship that is behind its cheaper siblings on an input type is an obvious gap.
• Web search reliability. Moonshot’s own documentation flags K3’s web_search as mid-update and does not currently recommend relying on it. That is the vendor’s caveat, not ours.
• An off switch for reasoning. K3 always reasons; reasoning_effort reaches low but never off. K2.6 offers a non-thinking mode. For high-volume mechanical work that difference is a real cost, and the flagship currently loses on it.
• Price, if the pattern holds. K3 costs 3.8× K2.6 per output token. Whether a successor narrows that or extends it is genuinely unknown.

We will not guess at context window sizes, parameter counts or prices. Those are the numbers people plan around, and inventing them is exactly what makes most coverage of unreleased models worthless.

Where does the Kimi K4 speculation come from?

From one report about graphics cards, not about a model. Secondary coverage attributes the story to The Information, published around 29 July 2026, reporting that Moonshot AI is seeking additional Nvidia Blackwell-class GPU capacity to train a next-generation model. We have not independently verified that report and are describing the attribution rather than vouching for its contents.

Read what that claim actually contains, and what it does not:

The reporting says	The reporting does not say
Moonshot is seeking more GPU capacity	That a model called K4 exists
The capacity is for a next-generation model	Any parameter count or context window
Sourced to a single publication, late July 2026	Any price, benchmark or release date

Buying compute is what a frontier lab does continuously. It is evidence of intent to train something, which was never in doubt, and it is evidence of nothing else.

What happened next is the more useful lesson. That single procurement story has been reworked downstream into headlines about what K4 “could be”, complete with speculative capability claims. None of it originates with Moonshot. If you find a page listing K4 specifications or pricing, you are looking at the end of that chain, not at a source.

How do you tell a real specification from an invented one?

Three checks, and they work for any unreleased model from any vendor — not just this one.

• Follow the citation to a primary source. A real specification traces to the vendor’s own documentation, model card, weight release or announcement. If the trail ends at another blog post, there is no trail.
• Look for a date on the number. A price or context window without a verification date is not information, particularly in a family where two models have been retired in four months.
• Distrust suspicious precision. “Approximately 2 million tokens” is the shape of an invented figure. Real published context windows are exact powers-of-two-ish numbers like 1,048,576, because they come from the implementation rather than from a writer’s sense of what sounds plausible.

That last one is worth internalising. Every context window in this family is a specific number Moonshot published, and we print them in full for exactly that reason.

What should you use today?

Kimi K3 if you need its 1,048,576-token context window or structured output. Otherwise Kimi K2.6 for general work or Kimi K2.7 Code for code — both cost considerably less per token and cover the large majority of workloads.

An unannounced model should have no bearing on what you build on now. What the release history should change is how you build: the two most recent retirements gave existing users a matter of months, so keep the model identifier in configuration rather than scattered through your code. Our model index shows the current lifecycle status of every model in the family.

How will this page change when something ships?

The same way every other page here does. Specifications and pricing go into our model registry with a link to Moonshot’s own documentation and the date we checked it, and every page that references those figures updates at once. Nothing gets published from a rumour, a screenshot or a repost.

If you want to know when that happens, the blog covers releases and retirements as they are announced, and there is an RSS feed. Until then, this page stays as it is: a record of what is known, which is currently nothing.