Kimi release history.

Moonshot AI has shipped 7 major model releases and 4 other major launches since January 2025. Kimi first launched as Kimi Chat in October 2023 with a 200,000-character context window; Moonshot AI’s own sources give no day for that launch or its 2024 releases, so this timeline starts with Kimi k1.5 in January 2025. The latest model recorded here is Kimi K3, announced 16 Jul 2026. Every entry is dated from Moonshot AI’s own announcement and links to it.

11 MAJOR RELEASES2025–2026CHECKED 23 SEP 2026
NEWEST KIMI K
Kimi K3 · 16 Jul 2026
NEWEST CODE
Kimi K2.7 Code · 12 Jun 2026
§ 01 — LINEAGE

How the Kimi models evolved.

KIMI KKimi K2 — 2025-07-112Kimi K2.5 — 2026-01-272.5Kimi K2.6 — 2026-04-202.6Kimi K3 — 2026-07-163CODEKimi K2.7 Code — 2026-06-12K2.7THINKINGKimi k1.5 — 2025-01-20k1.5Kimi K2 Thinking — 2025-11-06K2
CURRENTSUPERSEDEDRETIREDLARGER SQUARE = WATERSHED
§ 02 — MODELS

Every Kimi model release.

7 model releases, newest first. “Current” is the newest release in its line; older ones are superseded, and retired ones have left Moonshot AI’s apps or API. Where a model is in our benchmark tables, the last column links to its scores, price in AUD and Australian availability.

MODEL RELEASES7 RELEASES
Kimi model releases, newest first
RELEASEANNOUNCEDLINESTATUSBENCHMARKS
Kimi K3WATERSHEDKimi KCURRENTKimi K3 →
Kimi K2.7 CodeCodeCURRENT
Kimi K2.6Kimi KSUPERSEDED
Kimi K2.5Kimi KRETIRED
Kimi K2 ThinkingThinkingRETIRED
Kimi K2WATERSHEDKimi KRETIRED
Kimi k1.5ThinkingSUPERSEDED
§ 03 — TIMELINE

Every major release, year by year.

11 OF 11

20266 RELEASES

  1. MODEL

    Kimi K3

    • WATERSHED
    • LAUNCHED
    • CURRENT
    • K3 AND KIMI WORK

    Kimi K3 is a 2.8T-parameter model built on Kimi Delta Attention and Attention Residuals, with native vision and a 1-million-token context, available on Kimi.ai, the apps, Kimi Work, Kimi Code and the API. Full weights followed on 27 July 2026 under the Kimi K3 License.

    Why it mattered. Moonshot calls it the first open 3T-class model; it said K3 still trails Claude Fable 5 and GPT 5.6 Sol overall but outperformed the other models it tested.

    AVAILABLE TO
    Open weights · Web · App · Kimi Work · Kimi Code · API
    BENCHMARKS
    Kimi K3 →

  2. MODEL

    Kimi K2.7 Code

    • OPEN WEIGHTS
    • CURRENT
    • K3 AND KIMI WORK

    Kimi K2.7 Code, a coding-focused agentic model built on K2.6 (1T total and 32B active parameters, 256K context), is released as open weights under a modified MIT licence and on the Kimi API and Kimi Code, followed by a HighSpeed API variant.

    Why it mattered. It was Kimi’s first dedicated coding model, and Moonshot said it uses about 30% fewer thinking tokens than K2.6 while improving long-horizon coding.

    AVAILABLE TO
    Open weights · API · Kimi Code

  3. PRODUCT

    Kimi Work

    • BETA
    • K3 AND KIMI WORK

    Kimi Work launches in beta as a desktop app for Mac and Windows, with a Chat mode and a Work mode that acts as a local agent on the user’s files, in single-agent or Agent Swarm execution.

    Why it mattered. It took Kimi’s agents off the web and onto the desktop, where they can work directly with local files for everyday knowledge work.

    AVAILABLE TO
    Mac · Windows

  4. MODEL

    Kimi K2.6

    • OPEN WEIGHTS
    • SUPERSEDED
    • AGENT SWARM

    Kimi K2.6 launches as an open-weights model (1T total and 32B active parameters, 256K context, text, image and video input) aimed at long-horizon coding, with Agent Swarm scaled to 300 sub-agents and 4,000 coordinated steps. It is available on Kimi.ai, the Kimi app, the API and Kimi Code.

    Why it mattered. Moonshot showed K2.6 working through coding tasks for more than 12 hours and 4,000 tool calls, and K2.6 became the base for Kimi K2.7 Code.

    AVAILABLE TO
    Open weights · Web · App · API · Kimi Code

  5. MODEL

    Kimi K2.5

    • OPEN WEIGHTS
    • RETIRED
    • AGENT SWARM

    Kimi K2.5 is an open-weights, natively multimodal model (1T total and 32B active parameters, 256K context) with instant and thinking modes. The Kimi app gains K2.5 Instant, Thinking, Agent and Agent Swarm (beta) modes, and K2.5 reaches the API and Kimi Code.

    Why it mattered. Agent Swarm lets K2.5 direct up to 100 sub-agents across up to 1,500 tool calls, which Moonshot said cut execution time by up to 4.5x, and vision and thinking now share one model.

    31 Aug 2026: kimi-k2.5 was discontinued on the Kimi API; Moonshot recommends Kimi K3.

    AVAILABLE TO
    Open weights · Web · App · API · Kimi Code

  6. PRODUCT

    Kimi Code

    • LAUNCHED
    • AGENT SWARM

    Alongside K2.5, Moonshot introduces Kimi Code, an open-source coding agent that runs in the terminal and integrates with IDEs including VS Code, Cursor and Zed, with image and video input.

    Why it mattered. It gave Kimi its own coding-agent product, sold through Kimi membership plans, and later models including K2.7 Code and K3 launched in it.

    AVAILABLE TO
    Terminal · IDEs

20255 RELEASES

  1. MODEL

    Kimi K2 Thinking

    • OPEN WEIGHTS
    • RETIRED
    • OPEN K2 AGENTS

    Kimi K2 Thinking is an open-weights thinking agent that reasons step by step while calling tools, sustaining 200–300 sequential tool calls, with a 256K context and native INT4 quantisation. It launched in Kimi’s chat mode and on the API.

    Why it mattered. Moonshot reported 44.9% on Humanity’s Last Exam with tools and 60.2% on BrowseComp, and said INT4 roughly doubled generation speed.

    25 May 2026: kimi-k2-thinking and kimi-k2-thinking-turbo were discontinued on the Kimi API; Moonshot recommends Kimi K3.

    AVAILABLE TO
    Open weights · Web · API

  2. AGENT

    OK Computer agent mode

    • LAUNCHED
    • OPEN K2 AGENTS

    Kimi launches Agent mode, code-named OK Computer, which plans and carries out multi-step tasks on its own and delivers finished work such as websites, documents and data analysis.

    Why it mattered. It moved Kimi from answering questions to completing tasks, and later releases extended Agent mode with office file output and Agent Swarm.

    AVAILABLE TO
    Kimi

  3. MODEL

    Kimi K2

    • WATERSHED
    • OPEN WEIGHTS
    • RETIRED
    • OPEN K2 AGENTS

    Moonshot releases Kimi K2, a mixture-of-experts model with 1 trillion total and 32 billion active parameters and a 128K context, as open weights (Base and Instruct) under a modified MIT licence, free in the Kimi web and mobile apps and on the Kimi API.

    Why it mattered. K2 put a trillion-parameter model built for agentic tool use, not just chat, into open weights, and the later K2.5 was trained on top of K2-Base.

    25 May 2026: The kimi-k2 API models, including the 0711 and 0905 previews and Turbo, were discontinued; Moonshot recommends Kimi K3.

    AVAILABLE TO
    Open weights · Web · App · API

  4. AGENT

    Kimi-Researcher

    • GRADUAL ROLLOUT
    • REASONING AND RESEARCH

    Kimi-Researcher, an autonomous agent for multi-turn search and reasoning trained end to end with agentic reinforcement learning, begins a gradual rollout in Kimi with a waitlist. It works with search, web browsing and code execution tools.

    Why it mattered. It was Kimi’s deep research agent, and Moonshot reported 26.9% Pass@1 on Humanity’s Last Exam, up from 8.6% before reinforcement learning.

    AVAILABLE TO
    Kimi · Waitlist

  5. MODEL

    Kimi k1.5

    • ANNOUNCED
    • SUPERSEDED
    • REASONING AND RESEARCH

    Moonshot publishes Kimi k1.5, a multimodal model trained jointly on text and vision with reinforcement learning scaled to a 128K context, with its technical report on GitHub.

    Why it mattered. Moonshot claimed long chain-of-thought results matching OpenAI o1 across maths, coding and vision benchmarks, and short chain-of-thought results well ahead of GPT-4o and Claude 3.5 Sonnet.

    AVAILABLE TO
    Technical report
§ 04 — ERAS

4 eras of Kimi.

  1. ERA 01JANUARY 2025 – JUNE 2025

    Reasoning and research

    Kimi k1.5 brings reinforcement-learning reasoning to Kimi, and Kimi-Researcher starts rolling out as an autonomous research agent.

    2 releases, from Kimi k1.5 →
  2. ERA 02JULY 2025 – NOVEMBER 2025

    Open K2 agents

    Kimi K2 makes Moonshot an open-weights lab at trillion-parameter scale, followed by the OK Computer agent mode and K2 Thinking.

    3 releases, from Kimi K2 →
  3. ERA 03JANUARY 2026 – APRIL 2026

    Agent Swarm

    K2.5 folds vision, thinking and Agent Swarm into one open model alongside Kimi Code, and K2.6 scales swarms to 300 sub-agents.

    3 releases, from Kimi Code →
  4. ERA 04JUNE 2026 – NOW

    K3 and Kimi Work

    Kimi Work brings agents to the desktop, K2.7 Code adds a dedicated coding model, and K3 arrives as a 2.8T-parameter open model with a 1M-token context.

    3 releases, from Kimi Work →
§ 05 — QUESTIONS

Kimi releases, answered.

As of our 23 Sep 2026 review, Kimi K3, announced by Moonshot AI on 16 Jul 2026. Kimi K3 is a 2.8T-parameter model built on Kimi Delta Attention and Attention Residuals, with native vision and a 1-million-token context, available on Kimi.ai, the apps, Kimi Work, Kimi Code and the API. Full weights followed on 27 July 2026 under the Kimi K3 License.

Kimi first launched as Kimi Chat in October 2023 with a 200,000-character context window; Moonshot AI’s own sources give no day for that launch or its 2024 releases, so this timeline starts with Kimi k1.5 in January 2025.

As of 23 Sep 2026, the newest release in each line is: Kimi K2.7 Code (12 Jun 2026) and Kimi K3 (16 Jul 2026). Older models in those lines have been superseded, and some have been retired from Moonshot AI’s apps.

11 on this timeline: 7 model releases and 4 product, platform, agent and plan launches. It leaves out incremental updates, settings and feature drops — only releases that changed what Kimi could do or who could use it.

Kimi K3: it can be called from AWS Bedrock in Sydney and Melbourne, but only through global routing, so requests may be processed outside Australia; its weights are public, so it can also be self-hosted in Australia. Our benchmark pages track Australian cloud availability for each model weekly.

Dates are the day Moonshot AI announced each release, taken from Moonshot AI’s own site and documentation.

Major releases only: new models, products, plans and capability launches. Retirements are noted on the model they retired.

Checked 23 Sep 2026 · How the current models compare →

MAKE SENSE OF THE NEXT RELEASE

Choose AI improvements around your business needs.

DISCUSS YOUR AI OPTIONSAI SUPPORT & MAINTENANCE