Kimi K2

Why does this page exist if Kimi K2 is gone?

Because people still search for it, and much of what they find is wrong. Moonshot AI discontinued the entire kimi-k2 series on May 25, 2026 — including kimi-k2-thinking and the turbo variants — and those endpoints no longer respond. Plenty of pages still quote its API pricing as though you could buy it.

You cannot. If you have landed here from a comparison table, a tutorial or a pricing round-up that lists Kimi K2 as a live option, that page has not been updated since at least May 2026.

This is the clearest example we have of why every figure on this site carries a verification date. Moonshot retires models faster than most published material about them gets revised.

What should you use instead?

Kimi K3 is the current flagship, but it is probably not your answer. For most work that used to run on K2, Kimi K2.6 or Kimi K2.7 Code is the better landing place — both cost substantially less per token than K3 and both far exceed anything K2 documented.

Your K2 workload	Move to	Why
General text, chat, extraction	Kimi K2.6	Cheapest current model, multimodal, web search
Code	Kimi K2.7 Code	Same price as K2.6 on output, tuned for coding
Very large prompts	Kimi K3	1,048,576-token context window
Strict output schemas	Kimi K3	Structured output is K3-only
You need K2 specifically	Self-host	Weights are still published

The migration is usually trivial in code terms — the API is Chat Completions-shaped throughout, so in most cases you are changing a model identifier string. Our API quickstart covers the current setup.

What was Kimi K2, and what did it actually do?

It was Moonshot’s open-weights family from 2025, released July 11, 2025 with 1T parameters. Compared with anything currently shipping it was narrow: Moonshot documented text input and tool calling, and that is close to the whole list.

No image input, no video, no thinking mode, no JSON mode, no context caching, no web search. The capability matrix above is unusually sparse, and that sparseness is the point — it shows how much the family added in roughly a year.

Moonshot also never published a context window figure for K2 that we have been able to verify, which is why this page shows a dash rather than a number. We would rather print nothing than repeat a figure we cannot trace to Moonshot’s own documentation, and there are several different numbers circulating for this model.

Arrived here from an older tutorial?

Then three things in it are probably out of date, and it is worth checking all three before you debug anything. Code written against Kimi K2 does not fail informatively — you get an error about an unknown model, or nothing that points at the real cause.

• The model identifier. kimi-k2-0711-preview, kimi-k2-0905-preview, kimi-k2-turbo-preview, kimi-k2-thinking and kimi-k2-thinking-turbo are all dead, as are the older kimi-latest and kimi-thinking-preview aliases. Replace with {k26.apiId} for general work or {code.apiId} for code. The full list with dates is on the model index.
• The pricing. Any figure quoted for K2 is describing a product that no longer sells. Current rates are on each model page here, each with the date we checked it.
• The capability assumptions. K2 documented text and tool calling only. A tutorial built around that will not be using vision, context caching or thinking mode — all of which are available now and change how you would write the same thing today.

What were kimi-k2-thinking and the turbo variants?

Members of the same discontinued series, and they went at the same time. Moonshot retired the whole kimi-k2-* family on May 25, 2026 — the base model, kimi-k2-thinking, and the turbo variants alike. None of those identifiers resolve any more.

This matters because those names appear in a lot of 2025 and early-2026 material. Tutorials, comparison posts and starter templates from that period frequently hardcode kimi-k2-thinking in particular, since it was the reasoning option before reasoning became standard across the family. If you are working from any guide written before mid 2026, assume the model identifier in it is dead.

The capability it represented has not gone anywhere — it has become the default. Thinking mode is documented on Kimi K2.6, and Kimi K3 reasons on every request whether you ask it to or not. What was once a separate model is now a parameter, or in K3’s case not even optional.

So the migration from kimi-k2-thinking is not “find the new thinking model”. It is “pick a current model and use its thinking mode”, which for most workloads means K2.6.

Can you still run it yourself?

Yes. K2 was an open-weights release and the weights remain downloadable, so self-hosting still works exactly as it did before the API shutdown. What Moonshot retired was the hosted service, not the model.

Whether you should is a different question. K2 lacks capabilities that current models treat as baseline — no context caching, no vision, no thinking mode — so self-hosting it means running infrastructure to get a materially less capable model than the one you can call today for a published price. The defensible reasons are data residency, an existing fine-tune you cannot reproduce, or reproducing a historical result.

As with every open-weights release in this family, downloadable is not the same as open source. Read the licence attached to the weight release before deploying commercially or redistributing anything derived from it.

What does the retirement pattern tell you?

That Kimi model lifetimes are short, and you should plan for that rather than be surprised by it. K2 was discontinued in May 2026; Kimi K2.5 is scheduled to follow on 31 August 2026. That is two retirements inside four months.

Put a number on it and the pattern is sharper still. Kimi K2 was released July 11, 2025 and discontinued May 25, 2026 — a hosted lifetime of roughly ten months. Kimi K2.5 arrived in January 2026 and shuts down at the end of August 2026, which is about eight. Neither model reached its first birthday on the API.

And those are only the two with their own pages. Moonshot’s deprecation notices record four separate shutdown dates in under a year: kimi-thinking-preview in November 2025, kimi-latest in January 2026, the whole kimi-k2-* series in May 2026, and Kimi K2.5 at the end of August 2026. Every discontinued identifier is listed with its date on the model index.

That is the planning horizon you are actually working with, and it is shorter than most teams assume when they pick a model. It is not unique to Moonshot, but Moonshot moves faster than most, and the retirements have so far come with limited notice.

The practical response is architectural. Keep the model identifier in configuration rather than scattered through your code, so a migration is a deploy rather than a refactor. Watch the status of every model rather than assuming the one you integrated last year is still current. And treat any pricing or capability claim you read elsewhere as stale until you find a date on it.

We track these dates because nobody else in this niche seems to. Every model page here carries the date its figures were last checked against Moonshot’s own documentation, and we publish shutdowns as blog posts when they are announced rather than quietly editing a table.