Try Kimi at kimi.com →

Kimi AI Models
Versions and Differences

Moonshot AI has released a new Kimi model every few months, from Kimi K2 in July 2025 to the flagship Kimi K3.1 Pro in October 2026. This guide lists every current version, explains how they differ in size, context, speed and price, and helps you choose the right one.

6 current models Up to 1M-token context API from $0.95 / $4
THE KIMI MODEL FAMILY Bar height = total parameters · label = context window 1T 2T 3T 128K1T Kimi K2 Jul 2025 256K1T K2.5 Jan 2026 256K1T K2.6 Apr 2026 256K1T K2.7 Code Jun 2026 1M2.8T Kimi K3 Jul 2026 1M? K2.8 Preview Sep 2026 1M2.8T K3.1 Pro Oct 2026 NEWEST 2.8× bigger
LINEUP

Current Kimi models

Six Kimi models are in use today. Two are flagships built on the same 2.8T K3.1 Pro model, one is the previous flagship, one powers Kimi Code, and two are low-cost open models from the K2 line.

COMPARE

Kimi models side by side

The main specifications of every current Kimi model. API prices are in US dollars per million tokens. Scroll sideways on a phone.

FeatureK3.1 ProK3.1 Pro HighSpeedKimi K3K2.8 PreviewK2.7 CodeKimi K2.6
ReleasedOct 2026Oct 202616 Jul 202611 Sep 202612 Jun 202620 Apr 2026
Best forHardest workSpeedLegacy K3 appsKimi CodeCheap codingCheap all-round
Total parameters2.8T2.8T2.8TNot published1T1T
Active per token104B104B16 of 896 expertsNot published32B32B
Context window1M1M1M1M256K256K
InputsText, image, videoText, image, videoText, image, videoText, imageText, code, imageText, image
ThinkingLow / High / MaxLow / High / MaxLow / High / MaxLow / High / MaxAlways onThinking or instant
API input / output$2 / $10$4 / $20$3 / $15—$0.95 / $4$0.95 / $4
Cached input$0.20—$0.30—$0.19$0.16
Open weightsYesSame weightsYesNoYesYes

Sources: Moonshot AI launch posts, the Kimi API pricing pages and Kimi Code release notes. Moonshot has not published K2.8 Preview's size or a public API price for it.

DIFFERENCES

Key differences explained

The models differ in six ways that matter when you choose one.

Size

The K3 models have 2.8 trillion parameters, almost three times the 1 trillion of the K2 line. Bigger models usually reason better but cost more to run and host.

Context window

K3.1 Pro, K3 and K2.8 Preview read about 1 million tokens at once. K2.6 and K2.7 Code read 256K, which is still enough for long documents and mid-size codebases.

Thinking effort

The K3 models and K2.8 Preview let you choose Low, High or Max thinking. K2.7 Code always thinks, and K2.6 can answer instantly or think first.

Vision and video

All current models read images. The K3 models also understand video, which helps with screen recordings and visual bugs.

Price

API prices range from $0.95 / $4 for K2.6 and K2.7 Code to $3 / $15 for K3. K3.1 Pro sits in between at $2 / $10, and cached input costs about a tenth.

Open or closed

Almost every Kimi model ships open weights you can download and self-host. K2.8 Preview is the exception, available only in Kimi Code and Kimi Work.

TRADE-OFFS

Context window vs price

More reading, or a lower bill?

The newest models can read about four times more text at once, but the older K2 models cost much less per token. Pick based on what your task actually needs.

  • 1M
    Choose a 1M-context model (K3.1 Pro, K3, K2.8 Preview) for whole codebases, long contracts or many reports at once.
  • 256K
    Choose a 256K model (K2.6, K2.7 Code) when inputs fit and cost matters. 256K tokens is still a few hundred pages.
  • $
    Cache repeated input. Cached input costs about a tenth of normal input on every model, which matters most on the big ones.
Context window Context window Output price Output price HOW MUCH IT CAN READ AT ONCE API OUTPUT PRICE PER 1M TOKENS K3.1 Pro 1M $10 Kimi K3 1M $15 K2.8 Preview 1M Kimi Code only K2.7 Code 256K $4 Kimi K2.6 256K $4 The newest models read 4× more than K2.6 and K2.7 Code K2.6 and K2.7 Code cost less than half of K3.1 Pro

Context window and API output price for the five current models with published figures.

CHOOSE

Which Kimi model should I use?

Pick your job

Tap the job that sounds most like yours to see the model we'd start with.

Best model for thisKimi K3.1 Pro

The 1M-token context holds whole stacks of reports, and it made the fewest false statements in Moonshot's finance tests. Set thinking to Max for the hardest questions.

Read the guide →

Using the Kimi app? You don't pick a version there. Choose a mode, and Kimi uses its newest models.

PICK A MODEL BY THE JOB YOUR JOB BEST MODEL Kimi K3.1 Pro$2 / $10 · 1M context K2.8 PreviewKimi Code default K2.7 Code$0.95 / $4 · open K3.1 Pro HighSpeed2–3× faster replies Kimi K2.61T · open weights Hard research & analysis Waiting on every reply Daily coding in Kimi Code Cheap coding via the API Self-host on less hardware Hard research & analysis Kimi K3.1 Pro$2 / $10 · 1M context Waiting on every reply K3.1 Pro HighSpeed2–3× faster replies Daily coding in Kimi Code K2.8 PreviewKimi Code default Cheap coding via the API K2.7 Code$0.95 / $4 · open Self-host on less hardware Kimi K2.61T · open weights

A quick map from common jobs to the best Kimi model for each.

HISTORY

Kimi model version history

  1. Jan 2025Kimi k1.5A reasoning model that could read images, and the step before the K2 line.
  2. Jul 2025Kimi K2The first 1-trillion-parameter open-weight Kimi model, with 32B active parameters and a 128K context. An update in September 2025 raised the context to 256K.
  3. Nov 2025Kimi K2 ThinkingAdded step-by-step reasoning and long chains of tool calls to K2.
  4. Jan 2026Kimi K2.5Made Kimi natively multimodal and introduced Agent Swarm with up to 100 sub-agents. Now retired from the API.
  5. Apr 2026Kimi K2.6Stronger long-horizon coding and agent work, with larger agent swarms.
  6. Jun 2026K2.7 CodeA coding specialist using about 30% fewer reasoning tokens than K2.6. HighSpeed followed three days later.
  7. Jul 2026Kimi K3A jump to 2.8T parameters, a 1M-token context and video input. Open weights followed on 27 July.
  8. Sep 2026K2.8 PreviewA closed, efficient mid-tier model that became the Kimi Code default.
  9. Oct 2026Kimi K3.1 ProThe current flagship: cheaper than K3, more accurate and better at long inputs. Open weights on 4 November.
KIMI MODEL RELEASES Each step: more capable, longer memory Kimi K2Jul 2025 Kimi K2.5Jan 2026 Kimi K2.6Apr 2026 K2.7 CodeJun 2026 Kimi K3Jul 2026 K2.8 PreviewSep 2026 K3.1PRO Kimi K3.1 ProOct 2026 · flagship 2.8T params · 1M context Open weights on Hugging Face

From the original Kimi K2 in July 2025 to the flagship Kimi K3.1 Pro in October 2026.

IN THE APP

Which models Kimi's products use

ProductModelHow to choose
Kimi app and kimi.comNewest models, with K3.1 Pro for the hardest workPick a mode (chat, thinking, agent, Agent Swarm). Kimi picks the model.
Kimi CodeK2.8 Preview by default (kimi-for-coding)Switch to a K3 model for harder jobs. HighSpeed uses more quota.
Kimi WorkK2.8 Preview and the K3 modelsSet in the app for each task.
Kimi APIK3.1 Pro, K3, K2.7 Code, K2.6Set the model field in your request.
Open weightsK3.1 Pro, K3, K2.7 Code, K2.6, K2.5, K2Download from Hugging Face and run with vLLM or SGLang.

In Kimi Code, turning thinking off sends K3-series and K2.8 Preview requests to a non-thinking K2.8 Preview variant. See Kimi Code and plans and credits.

DEVELOPERS

Kimi API model IDs

Use these IDs in the model field. The API is OpenAI-compatible at https://api.moonshot.ai/v1, and GET /v1/models lists what your account can use.

ModelModel IDWhereCopy
Kimi K3.1 Prokimi-k3.1-proKimi API
Kimi K3kimi-k3Kimi API
K2.7 Codekimi-k2.7-codeKimi API
K2.7 Code HighSpeedkimi-k2.7-code-highspeedKimi API
Kimi K2.6kimi-k2.6Kimi API
K2.8 Previewkimi-for-codingKimi Code only

Moonshot has not published a separate API ID for K3.1 Pro HighSpeed. Check the model list on platform.kimi.ai. Full setup: Kimi API guide.

OPEN WEIGHTS

Open weights and licences

K3 line (2.8T)

  • Kimi K3.1 ProKimi K3.1 Pro License
  • Kimi K3Open weights since 27 Jul 2026
  • Download sizeAbout 1.56 TB (K3.1 Pro)
  • HardwareRoughly 8+ B300-class GPUs

K3.1 Pro is free for commercial use, with extra terms for resellers earning over $20M a year and for very large products.

K2 line (1T)

  • K2.7 CodeModified MIT License
  • Kimi K2.6Modified MIT License
  • K2.5 and K2Modified MIT License
  • HardwareAbout 8 H200-class GPUs (INT4)

The Modified MIT License allows commercial use. Very large products must show the Kimi name.

SPECIALISTS

Specialist models and agents

Besides the main chat models, Moonshot has released models and agents for specific jobs.

LEGACY

Older and retired models

ModelStatusMove to
Kimi K2.5Closed to new API users after K3, platform sunset 31 Aug 2026. Weights still on Hugging Face.K2.6 or K3.1 Pro
Kimi K2 Thinking / TurboDeprecated on the API.K2.6 or K2.7 Code
Kimi K2 (0711, 0905)Superseded. Weights still on Hugging Face.K2.6
Kimi k1.5Replaced by the K2 line.Any current model
Migrating old code? Most errors after an upgrade come from an old model ID. Change the model value to a current ID from the table above, and check whether the new model always thinks, since that changes token use.
FAQ

Kimi models FAQ

The current Kimi models are Kimi K3.1 Pro (the flagship), K3.1 Pro HighSpeed (a faster version of it), Kimi K3, K2.8 Preview (the default in Kimi Code), K2.7 Code and Kimi K2.6. Older versions such as Kimi K2 and K2.5 are retired from Moonshot's API, but their open weights can still be downloaded.

Kimi K3.1 Pro is the newest Kimi model. Moonshot AI launched it in October 2026 and published its open weights on Hugging Face on 4 November 2026. It has 2.8 trillion parameters and a 1,048,576-token context window.

Kimi K3.1 Pro is the most capable Kimi model, and it is also cheaper than Kimi K3 on the API. For everyday coding in Kimi Code, K2.8 Preview is the default. For the lowest API cost, K2.7 Code and K2.6 cost $0.95 per million input tokens and $4 per million output tokens.

Both have 2.8 trillion parameters and a 1M-token context. K3.1 Pro costs $2 / $10 per million tokens instead of $3 / $15, handles long inputs more efficiently with sparse attention, makes fewer false statements on finance tasks in Moonshot's tests, and adds stronger 3D, CAD and game-engine skills.

K2.8 Preview is a mid-tier coding and agent model that Moonshot rolled out in Kimi Code on 11 September 2026 under the existing kimi-for-coding model ID. Moonshot says it is close to K3 in quality but thinks more efficiently. It has a 1M-token context, but its weights and size have not been published.

Both are 1-trillion-parameter open-weight models with a 256K-token context and the same API price. K2.6 is the general-purpose model, released in April 2026. K2.7 Code, released in June 2026, is tuned for long coding sessions, and Moonshot says it uses about 30% fewer reasoning tokens than K2.6.

Kimi K3.1 Pro, K3, K2.7 Code, K2.6, K2.5 and K2 all have open weights on Hugging Face. The K2 models use a Modified MIT License, and K3.1 Pro has its own Kimi K3.1 Pro License. K2.8 Preview is the exception: it is only available inside Kimi's own products.

Kimi K2.5, released in January 2026, was closed to new API users after Kimi K3 launched and was scheduled to leave Moonshot's platform on 31 August 2026. Its open weights are still on Hugging Face, and Agent Swarm has moved on to the K3 models.

The Kimi app and website run on Moonshot's newest models, with K3.1 Pro behind the most demanding answers and agent tasks. Kimi Code uses K2.8 Preview by default. In Kimi Code, turning thinking off sends K3-series and K2.8 Preview requests to a non-thinking K2.8 Preview variant.

HighSpeed is the same model served on faster infrastructure, not a different model. K3.1 Pro HighSpeed is two to three times faster at $4 / $20 per million tokens. K2.7 Code HighSpeed (kimi-k2.7-code-highspeed) reaches up to about 260 tokens per second, according to Moonshot.

Kimi K2.6 and K2.7 Code are the cheapest current Kimi API models at $0.95 per million input tokens and $4 per million output tokens. K3.1 Pro costs $2 / $10 and Kimi K3 costs $3 / $15.

Try the newest Kimi model

Chat with Kimi for free, or build with the API from $0.95 per million input tokens.