Try Kimi at kimi.com →

Kimi K2
Moonshot's Open 1T Model Family

Kimi K2 put Moonshot AI on the map in July 2025: a 1-trillion-parameter open model that rivalled paid ones at coding and tool use. It grew into a family, from K2 Thinking to K2.8 Preview. Here's what each version did, what still works today and where K2 is heading.

1T · 32B active Open weights K2.6 free in Kimi chat
THE KIMI K2 FAMILY · JULY 2025 TO SEPTEMBER 2026 JUL 2025K21T open model Retired SEP 2025K2-0905256K context Retired NOV 2025K2 ThinkingReasoning agent Retired JAN 2026K2.5Vision + Swarm Retired APR 2026K2.6Open, agentic Free in chat JUN 2026K2.7 CodeCoding model Retiring SEP 2026K2.8 PreviewKimi Code Preview Status as of November 2026. Successors: Kimi K3 and K3.1 Pro.
OVERVIEW

Kimi K2 at a glance

1TParameters

About 32 billion active for each token.

15.5TTraining tokens

Trained with the MuonClip optimiser, with no loss spikes.

65.8%SWE-bench Verified

At launch, ahead of GPT-4.1's 54.6% in Moonshot's tests.

MITModified licence

Open weights you can download and run.

STATUS

What still works today

Kimi chat: K2.6 is free

In the Kimi app, K2.6 is free for everyone in chat and doesn't use credits.

API: closing down

K2.6 and K2.7 Code were closed to new API users on 31 October 2026 and retire by the end of December 2026.

Kimi Code: K2.8

K2.8 Preview runs Kimi Code's default kimi-for-coding ID, and K2.7 Code Highspeed is still offered.

Open weights: forever

K2, K2 Thinking, K2.6 and K2.7 Code weights stay on Hugging Face for self-hosting.

Building something new? Moonshot points API users to Kimi K3.1 Pro. See the Kimi updates timeline for every change.
FAMILY

Every Kimi K2 version

VersionReleasedWhat was newStatus, Nov 2026
Kimi K211 Jul 20251T MoE, 32B active, 128K context; Base and InstructRetired from the API
K2-Instruct-0905Sep 2025Better coding and front-end work; 256K contextRetired from the API
K2 Thinking6 Nov 2025Reasoning agent, 200–300 tool calls, native INT4Retired from the API
K2.527 Jan 2026Native vision, Agent Swarm, Kimi Code CLIRetired 31 Aug 2026
K2.620 Apr 2026Open-weight, stronger long coding and agents; 256KFree in Kimi chat; API retiring by end Dec 2026
K2.7 CodeJun 2026Coding specialist, about 30% fewer thinking tokensAPI retiring by end Dec 2026; Highspeed still in Kimi Code
K2.8 Preview11 Sep 2026Coding model close to K3, 1M contextCurrent in Kimi Code (kimi-for-coding)

Details: K2.7 Code, K2.8 Preview and all Kimi models.

SPECS

Original Kimi K2 specifications

SpecKimi K2 (July 2025)
ArchitectureMixture-of-Experts with Multi-head Latent Attention (MLA)
ParametersAbout 1 trillion total, 32 billion active per token
Experts384 experts, 8 chosen per token plus 1 shared
Layers / heads61 layers, 64 attention heads
Vocabulary160K tokens
Context128K tokens (256K from K2-0905 onwards)
Training data15.5 trillion tokens of web text, code, maths and knowledge
OptimiserMuonClip (Muon with QK-Clip)
LicenceModified MIT
Launch API price$0.60 input, $0.15 cached, $2.50 output per 1M tokens

From the Kimi K2 model card and technical report (arXiv 2507.20534).

TRAINING

How Kimi K2 was trained

  • 1
    MuonClip. Moonshot used the token-efficient Muon optimiser and added QK-Clip, which caps runaway attention values. The result was 15.5 trillion tokens of training without a single loss spike.
  • 2
    Sparse experts. 384 experts with only 8 active per token. Moonshot's scaling tests showed more sparsity gave better results for the same compute.
  • 3
    Synthetic agent data. A large pipeline generated tool-use tasks across many simulated tools and environments.
  • 4
    Joint reinforcement learning. K2 learned from both checkable tasks, like maths and code, and from its own critiques on open-ended ones.
TRAINING LOSSLower is better · illustration tokens trained → 15.5T Loss spikes (unstable)MuonClip: no spikes Moonshot reported zero loss spikes across 15.5T tokens. Curves are illustrative, not Moonshot’s data.

MuonClip kept K2's training smooth from start to finish.

BENCHMARKS

Kimi K2 benchmarks (July 2025)

TestWhat it measuresK2-Instruct
SWE-bench VerifiedFixing real GitHub issues (agentic, single attempt)65.8%
SWE-bench MultilingualSame, across languages47.3%
LiveCodeBench v6Competitive coding53.7%
Tau2-BenchTool use in customer-service tasks66.1
ACEBench (En)Tool and API calling76.5
AIME 2025Competition maths49.5
GPQA DiamondGraduate-level science75.1

Moonshot-reported, non-thinking mode. On 17 July 2025, K2 ranked as the top open model on the LMSYS Arena leaderboard. Our independent view: Kimi AI review.

THINKING

Kimi K2 Thinking (November 2025)

K2 Thinking turned K2 into a reasoning agent that thinks step by step and calls tools for hundreds of steps in a row. Moonshot reported state-of-the-art results at launch.

MeasureK2 Thinking
Humanity's Last Exam (with tools)44.9%
BrowseComp60.2%
SWE-bench Verified71.3%
Sequential tool calls200–300 without help
Context window256K tokens
PrecisionNative INT4, about 2× faster generation

Moonshot-reported, all at INT4 precision. Independent leaderboards later showed lower Humanity's Last Exam scores under different setups.

USE

How to use Kimi K2 today

Chat for free

Pick K2.6 in the Kimi app or on kimi.com. It's free in chat and doesn't use credits. How to use Kimi.

API (existing users)

kimi-k2.6 costs $0.95 input, $0.16 cached and $4 output per 1M tokens until it retires. Plan your move to K3.1 Pro.

Self-host

Download the weights from Hugging Face and serve them with vLLM or SGLang on a multi-GPU server.

import os from openai import OpenAI client = OpenAI(api_key=os.environ["MOONSHOT_API_KEY"], base_url="https://api.moonshot.ai/v1") # Existing API users only: K2 models closed to new users on 31 Oct 2026 reply = client.chat.completions.create( model="kimi-k2.6", messages=[{"role": "user", "content": "Summarise this changelog in 5 bullets: ..."}], ) print(reply.choices[0].message.content) # Moving before the end-of-December retirement? Change one line: # model="kimi-k3.1-pro"

Keys and billing: Kimi API guide.

COMPARE

Kimi K2 vs K3

Kimi K2 familyKimi K3 / K3.1 Pro
Parameters1T total, 32B active2.8T total, 104B active
Context128K–256K1M
VisionFrom K2.5Built in (images and video)
API price per 1M (in / out)$0.95 / $4 (K2.6)$2 / $10 (K3.1 Pro)
StatusRetiring from the API by end of Dec 2026Current

More on the successors: Kimi K3 and Kimi K3.1 Pro.

FAQ

Kimi K2 FAQ

Kimi K2 is a family of open-weight AI models from Moonshot AI. The original K2, released on 11 July 2025, is a Mixture-of-Experts model with about 1 trillion parameters, 32 billion active per token. Later versions include K2 Thinking, K2.5, K2.6, K2.7 Code and K2.8 Preview.

The open weights are free to download under a Modified MIT licence. In the Kimi app, K2.6 is free to use in chat and doesn't use credits. API use is paid per token.

On the Kimi API, yes. K2-series models such as K2.6 and K2.7 Code were closed to new users on 31 October 2026 and are due to retire by the end of December 2026. K2.5 and the original K2 IDs are already retired.

The original K2 answers directly without long reasoning. K2 Thinking, released in November 2025, thinks step by step and can make 200–300 tool calls in a row, which made it much stronger on research and coding agent tasks.

At launch it scored 65.8% on SWE-bench Verified and ranked as the top open model on the LMSYS Arena (17 July 2025). K2 Thinking later scored 44.9% on Humanity's Last Exam with tools and 60.2% on BrowseComp.

MuonClip is the optimiser Moonshot used to train K2. It adds a step called QK-Clip to the Muon optimiser that stops attention values from growing too large, which let K2 train on 15.5 trillion tokens without a single loss spike.

For new projects, use K3.1 Pro: it's much stronger, has a 1M-token context and isn't retiring. K2.6 is still handy as a free chat model in the Kimi app, and the K2 open weights remain useful for self-hosting.

Yes, the weights are on Hugging Face (for example moonshotai/Kimi-K2-Instruct and moonshotai/Kimi-K2-Thinking). You'll need a multi-GPU server; the INT4 K2 Thinking weights are about 594 GB.

Try Kimi for free

K2.6 is free in Kimi chat, and K3.1 Pro has limited free use.