Kimi K2: 1T Open Model, API Set to Retire in December

In brief

Kimi K2 is Moonshot AI's open-weight model family, first posted to Hugging Face on 11 July 2025 with 1 trillion parameters, 32 billion active. As of 15 November 2026 it is no longer current: Moonshot lists six K2 API models as deprecated, has closed three more to new accounts, and plans to retire them by the end of December.

What is Kimi K2, and is it still Moonshot's current model?

Kimi K2 is a mixture-of-experts model: 384 experts, 8 picked for each token, plus 1 shared one. Moonshot trained it for agent work, meaning tool calls and code, and its GitHub page calls the first release "a reflex-grade model without long thinking." It isn't the current model. Kimi K3 took that slot in 2026, and Moonshot's API page says "We plan to retire all K2 series models by the end of December 2026".

A brushed silver crescent moon on a tall black plinth at the right of a charcoal studio, an open brass padlock lying on the floor beside it, and five identical grey stone tablets in descending height on a low black bench at the left.

Weight dates in the chart are when each repository was created on Hugging Face, read from its public API on 15 November 2026, and a repository can predate its public launch. The two API dates are Moonshot's. K2.5, a native multimodal model that Moonshot continued from K2 on about 15 trillion mixed text and image tokens, came between K2 Thinking and K2.6. Its repository date is left off because the post that announced it carries no date.

Kimi K2 shipped in July 2025, and its API is due to close in December 2026
Kimi K3 era, from 13 Jun 20262026K2weightsK2 InstructK2 0905K2 ThinkingK2.6K2.7 CodeK3weightsK3K3.1 ProK2 APIClosed to new accountsRetirement planned
K2 modelWhat it isWeights on Hugging FaceAPI status, read 15 Nov 2026
Kimi-K2-BaseFoundation model for fine-tuningYesNot on Moonshot's API model list
Kimi-K2-InstructThe first chat model, no long thinking, 128K contextYesDeprecated (kimi-k2-0711-preview)
Kimi-K2-Instruct-0905Update that Moonshot says supports 256K contextYesDeprecated (kimi-k2-0905-preview)
Kimi-K2-ThinkingReasoning model, native INT4, 256K contextYesDeprecated (kimi-k2-thinking and turbo)
Kimi-K2.5Native multimodal, images and video inYesDeprecated
Kimi-K2.6Multimodal agentic model for long coding jobsYesClosed to new accounts
Kimi-K2.7-CodeBuilt on K2.6, about 30% fewer thinking tokensYesClosed to new accounts, with a highspeed twin

Moonshot's two pages disagree by a day on the cutoff. The pricing page says K2 models are closed to new users "starting October 31, 2026". The model list says users registered before 1 November can keep K2.6 and K2.7 Code, with deprecation "expected in late December 2026". If you are not sure which one applies to you, treat 31 October as the day that it stopped.

Is Kimi K2 open source?

Kimi K2 is open weight under what Moonshot calls the Modified MIT License. The GitHub page says "Both the code and the model weights are released under the Modified MIT License." The license file, which carries the line "Copyright (c) 2025 Moonshot AI", is the MIT text plus one rule. A product or service built on K2 with more than 100 million monthly active users, or more than 20 million US dollars a month in revenue, "shall prominently display" Kimi K2 on its user interface.

Below those lines it behaves like plain MIT. You can download the weights, run them on your own machines, fine-tune them, and sell what you build with them. Moonshot's K2.6 and K2.7 Code cards carry the same Modified MIT wording. Kimi K3 ships under a different Kimi K3 License. Kimi vs ChatGPT covers that one, and open source LLMs guide covers what running weights like these takes.

The K2 GitHub page names four engines for hosting it: vLLM, SGLang, KTransformers and TensorRT-LLM. Its checkpoints are in block-fp8 format, and it recommends a temperature of 0.6. K2 Thinking went further and was trained for INT4, which Moonshot says roughly doubles generation speed. Even so, a model with 1 trillion parameters is a job for a data center. Not many of us will run it at home, and that is fine if all you need is the answers.

How good was Kimi K2 next to the models that followed it?

Each step up the K2 line raised the scores, and K3 raised them most. Every number below is Moonshot's own, copied from the model card for each release, so each was measured by Moonshot with its own settings and tools. The columns are the three tests that appear on more than one card.

Model, vendor-reportedHumanity's Last Exam with toolsBrowseCompSWE-bench Verified
Kimi K2 Instruct4.7 (text, no tools)Not reported65.8 (agentic, single attempt)
Kimi K2 Thinking44.9 (text)60.271.3
Kimi K2.551.8 (text)Not on the postNot on the post
Kimi K2.654.0 (full set)83.280.2
Kimi K356.0 (full set)91.2Not reported

Read the first column with care. K2 Thinking and K2.5 were scored on the text-only part of Humanity's Last Exam. K2.6 and K3 were scored on the full set. So that jump from 51.8 to 54.0 isn't a clean step, and you shouldn't read it as one. Kimi K3's card gives a different coding test instead: 67.5 on DeepSWE and 88.3 on Terminal-Bench 2.1.

The first K2 was strong on maths for a model with no long thinking. Its card lists 97.4 on MATH-500 and 75.1 on GPQA-Diamond, but only 53.7 on LiveCodeBench v6. Moonshot says K2 Thinking keeps tool use stable across 200 to 300 sequential calls, where it says earlier models degrade after 30 to 50 steps. That is Moonshot's claim from its card, and the card shows no separate data for it.

Can you still use the Kimi K2 API?

Only if you registered before the cutoff (31 October or 1 November, depending on which Moonshot page you read), and only for K2.6, K2.7 Code and its highspeed twin. Moonshot's model list shows the other six K2 model IDs as deprecated. The retirement notice covers the hosted API. It says nothing about the weights, and the Hugging Face repositories still answered on 15 November 2026.

If you run K2 code today, the move is to K3.1 Pro, Moonshot's newest model, which it calls "our most capable 3T-class open-weight model to date." The Kimi K3.1 Pro quickstart gives the settings, and the steps below follow it.

  1. Change the model ID to kimi-k3.1-pro.
  2. Point your OpenAI-compatible client at https://api.moonshot.ai/v1.
  3. Remove temperature, top_p, n, presence_penalty and frequency_penalty from requests, because K3.1 Pro fixes them.
  4. Set reasoning_effort to low, high or max. The default is max, and thinking can't be turned off.
  5. In multi-turn chats, send the whole assistant message back, not just its text.

That quickstart is the page that Moonshot links for K2 migration, so you should test your prompts against it before the cutoff. Step 4 is where bills change, since thinking is always on. Kimi pricing has the per-token rates Whizi read from Moonshot's list on 8 November, and is Kimi free covers the app.

Kimi K2 or Kimi K3: which one should you use?

Use K3 unless you are hosting K2's weights yourself. K2 was a trillion-parameter model with no vision and no long thinking, and K3 is a bigger multimodal reasoning model with eight times the context. Moonshot's K3 card puts K3 at 2.8 trillion total parameters and 104 billion active, with 16 of 896 experts picked per token.

As of 15 Nov 2026Kimi K2 (original)Kimi K3
Total parameters1 trillion2.8 trillion
Active per token32 billion104 billion
Context window128K tokens1,048,576 tokens
Images and videoNo, per Moonshot's K2 postNative visual understanding
ThinkingNone in the first releaseBuilt for deep reasoning
LicenseModified MITKimi K3 License
Moonshot APIDeprecatedAvailable, with K3.1 Pro above it

K2 still wins in two places. Its Modified MIT license is the lighter of the two, and a model with 32 billion active parameters is far cheaper to serve than one with 104 billion. If you have a K2 deployment that works, nothing forces you to move the weights. A new API customer has no K2 model to buy.

Where can you chat with Kimi K2 without hosting it?

Whizi lists Kimi K2 on its Powerhouse plan, and Kimi K3 and K2.5 on Pro. The base K2 row, K2 Thinking and K2.7 Code need Powerhouse at $49.99 a month. Pro at $29.99 reaches K3 and K2.5, and its first week is $1.99. Each Kimi row costs 10 credits a message, and Starter and the free plan reach none of them.

It is not Moonshot's own app, and a row can leave a picker when its provider retires the model. If you only want K2 for the open weights, Whizi doesn't help, because you can't download from it. The Kimi in Whizi docs list every row, and the Kimi K2 model page has its context window and credit cost.

Workflow checklist
  • K2 means the July 2025 family of open 1 trillion parameter models, not Moonshot's current one
  • Six K2 API models are deprecated and three more are closed to accounts registered after 31 October 2026
  • The weights stay public on Hugging Face under the Modified MIT License
  • New API work should target kimi-k3.1-pro at https://api.moonshot.ai/v1
  • Scores on the model cards are Moonshot's own, so compare them only across rows that share a test
Common questions

Frequently asked questions

Is Kimi K2 open source?

Kimi K2 is open weight, not fully unrestricted. Moonshot releases the code and weights under a Modified MIT License, which is MIT plus a rule that products with over 100 million monthly active users or 20 million US dollars in monthly revenue must prominently display Kimi K2. Below that you can download, run and sell what you build. Open source LLMs guide explains the difference.

Is Kimi K2 still available?

Partly. As of 15 November 2026, Moonshot's API lists six K2 models as deprecated and has closed K2.6 and K2.7 Code to accounts registered after 31 October 2026, with retirement planned by the end of December. The weights remain on Hugging Face. Kimi pricing shows what replaced K2 on the price list.

How many parameters does Kimi K2 have?

Kimi K2 has 1 trillion total parameters and 32 billion active for each token, according to its Hugging Face card. It is a mixture-of-experts model with 384 experts, 8 selected per token and 1 shared. Kimi K3 is larger, at 2.8 trillion total and 104 billion active. Kimi vs DeepSeek sets the Kimi line against DeepSeek.

What is Kimi K2 best at?

Agent work: calling tools and writing code. Moonshot's card for the first K2 lists 65.8 on SWE-bench Verified in agentic mode with one attempt, and 97.4 on MATH-500. It had no vision and no long thinking, which later K2 releases added. For a comparison against ChatGPT, see Kimi vs ChatGPT.