Kimi (chatbot)
Last reviewed
Sources
34 citations
Review status
Source-backed
Revision
v10 · 5,582 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
34 citations
Review status
Source-backed
Revision
v10 · 5,582 words
Add missing citations, update stale details, or suggest a clearer explanation.
Kimi is an artificial intelligence chatbot and assistant developed by Moonshot AI, a Beijing-based Chinese AI startup, first released to the public on November 16, 2023.[2] Kimi distinguished itself from competitors through an industry-leading long context window, which initially supported 128,000 tokens (about 200,000 Chinese characters) and was later expanded to handle over 2 million Chinese characters in a single prompt.[3][16] Available as a web app at kimi.com and as iOS and Android apps, Kimi grew into one of China's most widely used AI assistants, reaching over 36 million monthly active users at its peak around October 2024.[17] It has powered a rapid model lineage including Kimi K1.5, Kimi K2, Kimi K2 Thinking, Kimi K2.5, and the April 2026 flagship Kimi K2 K2.6, which scaled its Agent Swarm to 300 sub-agents and matched or led closed frontier models such as GPT-5.4 and Claude Opus 4.6 on several public benchmarks.[23] By early 2026, Moonshot AI was reportedly raising fresh capital at an $18 billion valuation and weighing a Hong Kong listing.[21][28]
Moonshot AI (formally known as Dark Side of the Moon, or "Yuemian" in Chinese) was co-founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin.[1] All three founders were classmates at Tsinghua University and had been bandmates in a group called Splay. The company's name was inspired by Pink Floyd's album The Dark Side of the Moon, and Moonshot AI launched on the 50th anniversary of that album's release.
Yang Zhilin serves as CEO. He was born in 1992 and graduated first in his class from Tsinghua's Computer Science Department in 2015.[11] He then pursued a PhD at Carnegie Mellon University (CMU), studying under Ruslan Salakhutdinov (later Apple's director of AI research) and William Cohen (a principal scientist at Google DeepMind). Yang completed his doctorate in just four years, finishing in 2019.[11] During his time at CMU, he worked at Google Brain and Meta Platforms, and co-authored influential papers on Transformer-XL and XLNet, both of which advanced the state of the art in handling long sequences in language models.[11] In 2016, while still a student, Yang co-founded Recurrent AI, a company focused on applying Transformer-XL to sales technology.[10] After completing his PhD, he worked with Huawei Technologies on an early version of their Pangu AI model in 2020 and led development of the Wudao large language model at the Beijing Academy of Artificial Intelligence in 2021.[11]
Zhou Xinyu serves as CTO, and Wu Yuxin serves as COO. By late 2025, Moonshot AI employed roughly 300 people, up from approximately 40 at the time of its seed funding round.
Moonshot AI has raised substantial venture capital since its founding, becoming one of the best-funded AI startups in China.
| Round | Date | Amount | Lead Investor(s) | Valuation | Key Participants |
|---|---|---|---|---|---|
| Seed | June 2023 | ~$200M | HongShan (Sequoia China) | $300M | ZhenFund |
| Series B | February 2024 | $1B[3][4] | Alibaba, HongShan | $2.5B[3] | Meituan, Xiaohongshu |
| Series C | August 2024 | $300M | Tencent, Gaorong Capital | $3.3B | Existing investors |
| Series D | December 2025 | $500M[5] | Undisclosed | $4.3B | Existing investors |
| Series E (Tranche 1) | February 2026 | $700M+[6] | Alibaba, Tencent, 5Y Capital | $10B | IDG Capital |
| Series E (Tranche 2) | March 2026 | Up to $1B | Existing investors | $18B[21][27] | Alibaba, Tencent, IDG Capital, 5Y Capital |
Bloomberg reported on March 14, 2026 that Moonshot AI had secured commitments at the $18 billion valuation, more than quadrupling its valuation in roughly three months.[27] Total funding raised through early 2026 exceeds $2.7 billion, and the company has become one of the fastest Chinese startups to reach decacorn status (a valuation exceeding $10 billion).[7]
In late March 2026, Bloomberg and the South China Morning Post reported that Moonshot AI had begun discussions with China International Capital Corp. (CICC) and Goldman Sachs about a possible Hong Kong listing.[28] The deliberations followed Hong Kong IPOs by domestic rivals Zhipu AI and MiniMax in January 2026, which drew strong demand from mainland and overseas funds and reset investor appetite for Chinese AI equities. Yang Zhilin had publicly stated in late 2025 that Moonshot was "not in a rush" to go public, but the rapid valuation jump, the open-source success of Kimi K2 and K2.5, and a friendlier listing window appear to have shifted the calculus. As of April 2026, the company had not formally filed listing documents, and people familiar with the talks cautioned that internal deliberations could still result in no near-term IPO.[28]
Yang Zhilin met with Chinese Premier Li Qiang in early 2026 as part of Beijing's push to accelerate the digital transformation of manufacturing.[30] Li urged faster industrial AI adoption and held up Moonshot's work as an example of homegrown frontier model development.[30] The meeting placed Moonshot AI alongside DeepSeek, Unitree, and a handful of other firms that Beijing has highlighted as national AI champions. Yang was also the only Chinese independent large language model founder invited to speak in person at NVIDIA GTC 2026, where he described Kimi K2.5 as organized around "three bets": training efficiency, long-context capability, and multi-agent coordination.
Moonshot AI first introduced Kimi in October 2023 as an AI assistant capable of processing up to 200,000 Chinese characters per conversation, which corresponded to roughly 128,000 tokens.[2] This made Kimi the first publicly available Chinese AI model to support lossless context windows of that size. At the time, most competing chatbots in China and globally were limited to context windows of 4,000 to 32,000 tokens.
On November 16, 2023, Kimi was released to the general public through its web interface at kimi.com.[2] The chatbot was built on Moonshot AI's proprietary large language model and supported both Chinese and English.
Yang Zhilin had outlined three key milestones for the company's technology roadmap: long context length, a multimodal world model, and a scalable general architecture capable of continuous self-improvement without human input. The long-context capability was the first of these milestones, and it became Kimi's defining feature in the Chinese AI market. Yang has summarized the bet bluntly, telling one interviewer that "lossless long context is everything," and arguing that "every problem is essentially a long-context problem."[33]
In March 2024, Moonshot AI announced that Kimi could now handle 2 million Chinese characters in a single prompt, a tenfold increase from the original 200,000-character limit.[16] This upgrade, initially available in closed beta for invited users, generated enormous public interest.[16] The announcement went viral across Chinese social media and tech communities.
The surge of new users overwhelmed Kimi's infrastructure. On March 21, 2024, both the Kimi app and website suffered a prolonged outage lasting roughly two days due to server overload. Moonshot AI issued a public apology and worked to expand capacity to meet demand. The incident, while embarrassing, also underscored the strength of consumer interest in Kimi's long-context capabilities and marked a turning point for the product's public profile.
In July 2024, Kimi introduced a "context caching" feature in public beta. This optimization allowed the model to store and reuse portions of previously processed context across conversations, reducing latency and computational cost for users who repeatedly worked with the same large documents.
Kimi offers a broad set of features designed around its core strength of long-context understanding.
Kimi supports uploading and analyzing a wide range of file formats, including PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), TXT, and Markdown documents, as well as images (PNG, JPG, GIF, WebP up to 20MB each). Individual files can be up to 100MB, and users can upload and process up to 50 files simultaneously. The system can extract key information from uploaded documents, generate summaries, create comparison tables, and answer questions based on document content.
Kimi for Docs is a specialized feature that helps users write, convert, review, and translate documents. Users can upload PDFs, Word files, slides, or spreadsheets and receive well-structured outputs with summaries, tables, and polished formatting.
Kimi can perform real-time searches across more than 100 websites, retrieving up-to-date information on topics ranging from market trends to breaking news. The web search capability is tightly integrated with the chatbot's reasoning, allowing it to verify facts, synthesize information from multiple sources, and provide citations for its claims.
The extended context window allows Kimi to produce and maintain coherence across exceptionally long written outputs. Users can request essays, reports, articles, and other long-form content while maintaining consistency in tone, terminology, and argumentation across thousands of words.
Kimi supports code generation, debugging, and explanation across multiple programming languages. The Kimi Code (also known as KFC) product is a dedicated developer tool that functions as an AI agent in the terminal. Kimi Code CLI can read and edit code, execute shell commands, search and fetch web pages, and autonomously plan and adjust actions during execution. It integrates with mainstream IDEs such as VS Code and Zed, supporting multi-file refactoring, debugging, and automation at scale.
On October 11, 2024, Moonshot AI launched the Kimi Explore Edition, which added AI-powered autonomous search capabilities. Three notable enhancements were introduced:
On September 24, 2025, Moonshot AI announced "OK Computer," an agentic AI feature named after the Radiohead album.[12] This feature transforms Kimi into a self-sufficient agent with built-in computing power. When using OK Computer, users give simple instructions, and Kimi independently operates a virtual computer to perform complex tasks such as website development, data analysis, image and video generation, and presentation creation.[12] The feature supports processing up to 1 million rows of input data at once and can produce multimedia outputs including text, audio, and video. OK Computer was initially available on Kimi's website for international and Hong Kong users, with beta testing for mainland China users.[12]
Kimi Claw is a feature that lets users build 24/7 AI assistants with long-term memory and personality. These assistants can proactively execute scheduled tasks, acting as persistent automated agents rather than one-off conversational tools.
In June 2025, Moonshot AI released Kimi-Researcher, an autonomous AI research agent. Kimi-Researcher takes an average of 23 reasoning steps per task, visits over 200 websites, and solves problems through active exploration rather than following fixed instructions. The feature is designed for deep research tasks that require synthesizing information from many sources.
Kimi is accessible through multiple platforms:
| Platform | Details |
|---|---|
| Web | kimi.com, full-featured web application |
| iOS | Free download on the Apple App Store, listed as "Kimi - OK Computer is Live" |
| Android | Free download on the Google Play Store |
| PWA | Progressive Web App for Windows, macOS, and Linux |
| API | Developer API accessible through platform.moonshot.ai |
Conversation history, preferences, and subscriptions sync across all devices. The mobile app includes Siri Shortcuts integration on iOS, allowing users to trigger AI tasks through voice commands.
Kimi offers a free tier (called "Adagio") with generous functionality and no credit card requirement. Paid subscription tiers include:
| Plan | Monthly Price (China) | Monthly Price (International) | Features |
|---|---|---|---|
| Adagio (Free) | Free | Free | Basic access to all modes |
| Andante (Basic) | ~49 CNY | ~$19 USD | Moderate allotment of advanced features |
| Moderato (Premium) | ~99 CNY | ~$39 USD | Larger usage limits, faster response speeds |
In China, Kimi also offers short-term plans ranging from 5.2 yuan for four days to 399 yuan for a full year of priority access.
On January 20, 2025, Moonshot AI released Kimi K1.5, a reasoning model that introduced reinforcement learning-based reasoning capabilities. The accompanying technical paper, "Kimi k1.5: Scaling Reinforcement Learning with LLMs" (published on arXiv), detailed several key innovations:[8]
Moonshot AI claimed that K1.5's short-CoT performance outperformed GPT-4o and Claude 3.5 Sonnet on benchmarks like AIME and LiveCodeBench by up to 550%, while its long-CoT performance matched OpenAI o1 across multiple modalities including mathematics, coding, and multimodal reasoning.[8]
In July 2025, Moonshot AI released Kimi K2, a large language model using a mixture-of-experts (MoE) architecture.[9] Key specifications:
| Specification | Value |
|---|---|
| Total parameters | 1.04 trillion[9] |
| Active parameters per request | 32 billion[9] |
| Number of experts | 384 |
| Training data | 15.5 trillion tokens[9] |
| Attention mechanism | Multi-head Latent Attention (MLA) |
| Hidden dimension | 7,168 |
| Context window | 128K tokens (later expanded to 256K) |
| License | Modified MIT |
Kimi K2 was released as an open-source model, with weights made freely available for download and modification.[15] The model was trained using the MuonClip optimizer, an improvement upon Muon that incorporates a QK-clip technique to address training instability.[9]
On the LMSYS Arena leaderboard (as of July 17, 2025), Kimi K2 ranked as the top-performing open-source model and 5th overall, based on over 3,000 user votes. Notable benchmark results included 65.8 on SWE-bench Verified, 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, and 75.1 on GPQA-Diamond.[9]
On September 9, 2025, Moonshot released an updated version (Kimi-K2-Instruct-0905) that improved agentic coding performance and doubled the context window from 128K to 256K tokens.
In November 2025, Moonshot released Kimi K2 Thinking, an open-source update designed for advanced reasoning and agentic tasks. The model retained the 1-trillion-parameter MoE architecture with 32 billion active parameters and supported up to 256,000-token contexts.[14] Notably, the model was reportedly trained for approximately $4.6 million, a figure that attracted attention for its relative efficiency compared to the training costs of competing frontier models.[14] Moonshot researchers, however, cautioned that the $4.6 million figure was not an official number, telling CNBC that "it is hard to quantify the training cost because a major part is research and experiments."[34]
Kimi K2 Thinking could execute 200 to 300 sequential tool calls autonomously and used native INT4 quantization for efficiency, making it more practical to deploy.[14] Shipping natively in INT4 (rather than the FP8 precision used by earlier K2 Instruct releases) cut the model footprint to roughly 594GB, versus over 1TB for K2 Instruct, while roughly doubling generation speed.[34]
On January 27, 2026, Moonshot AI released Kimi K2.5, the most capable model in the Kimi series to date.[22] K2.5 introduced native multimodal vision capabilities through a 400-million-parameter vision encoder called MoonViT.[20] The model was trained on 15 trillion tokens mixing visual and textual data from the start, meaning vision and language capabilities developed together rather than being added separately.[20]
The model supports four operational modes on kimi.com and the Kimi app:[13]
| Mode | Description |
|---|---|
| K2.5 Instant | Fast responses for straightforward queries |
| K2.5 Thinking | Extended reasoning with chain-of-thought for complex problems |
| K2.5 Agent | Single-agent task execution with tool use |
| K2.5 Agent Swarm (Beta) | Parallel multi-agent coordination for complex workflows |
The Agent Swarm feature is the most distinctive capability of K2.5.[22] Trained using Parallel-Agent Reinforcement Learning (PARL), the model learns to self-direct a swarm of up to 100 sub-agents executing parallel workflows across up to 1,500 coordinated steps.[20] The orchestrator decomposes tasks into parallelizable chunks (such as finding sources, extracting data, verifying claims, and formatting output) and instantiates specialized sub-agents to execute each subtask.
This parallel approach reduces execution time by 4.5x compared to sequential processing.[20] On Humanity's Last Exam, K2.5 achieved 50.2% at 76% lower cost than Claude Opus 4.5.[20] The model also set state-of-the-art results for open-source models on VideoMMMU (65.2) and achieved 71.9 on Video-MME for general video understanding.[20] K2.5 also reached 74.9% on BrowseComp in standard agent mode and 78.4% with Agent Swarm enabled.[20]
On April 20, 2026, Moonshot AI released and open-sourced Kimi K2.6, the most capable model in the series at the time of writing.[23][25] K2.6 retained the trillion-parameter mixture-of-experts architecture (with 32 billion parameters activated per token) but emphasized stability and long-horizon execution rather than raw size.[24] The official tech blog framed K2.6 around "state-of-the-art coding, long-horizon execution, and agent swarm capabilities."[25]
| Specification | Value |
|---|---|
| Total parameters | ~1 trillion[24] |
| Active parameters per token | 32 billion[24] |
| Number of experts | 384 (8 selected per token, plus 1 shared expert always active) |
| Layers | 61 (one dense) |
| Attention mechanism | Multi-head Latent Attention (MLA), 64 heads |
| Attention hidden dimension | 7,168 |
| MoE hidden dimension per expert | 2,048 |
| Activation function | SwiGLU |
| Vocabulary | 160K tokens |
| Context window | 256K tokens (with stability improvements over K2.5) |
| Vision encoder | MoonViT, 400M parameters (image and video input) |
| Training tokens | ~15.5 trillion |
| License | Modified MIT |
Architecturally, K2.6 inherits the ultra-sparse MoE design used by K2 and K2.5, including the MuonClip optimizer that Moonshot developed to prevent the loss spikes that historically destabilize trillion-parameter MoE training.[24] The official communication stresses that the headline change versus K2.5 is not raw context size (256K was already supported in K2.5) but stability across that context, particularly during long-horizon coding sessions where the context window fills up over hours.[25]
The most heavily promoted capability of K2.6 is long-horizon execution, which Moonshot describes as the ability to keep working productively on a single task across many hours and thousands of tool calls.[23] Two demonstrations published with the launch:
| Task | Duration | Tool Calls | Outcome |
|---|---|---|---|
| Optimize Qwen3.5-0.8B inference in Zig | 12+ hours | 4,000+ | Throughput improved from ~15 to 193 tokens/second |
| Overhaul an 8-year-old financial matching engine (exchange-core) | 13 hours | 1,000+ | 185% median throughput gain, 133% performance improvement |
K2.6 scales the Agent Swarm framework to 300 sub-agents executing across up to 4,000 coordinated steps simultaneously, roughly tripling the K2.5 ceiling of 100 sub-agents and 1,500 steps.[23] The orchestrator continues to decompose heterogeneous tasks into parallel subtasks, but the larger swarm allows it to fan out across more domains at once.
K2.6 introduces a "Skills" feature that turns high-quality reference documents into reusable capabilities.[25] Users can upload PDFs, spreadsheets, slides, or Word files, and Kimi extracts the structural and stylistic patterns of the document. The captured Skill can then be reapplied to future tasks so that the swarm produces output in the same format and quality as the example. Demonstrations released with the launch included matching one uploaded CV against 100 California job postings to produce 100 tailored resumes; identifying 30 Los Angeles retail stores without websites from Google Maps and generating a landing page for each; and turning an astrophysics paper into a reusable academic Skill that produced a 40-page, 7,000-word follow-up paper alongside a 20,000-row dataset and 14 charts.[25]
Moonshot's launch materials and third-party reporting positioned K2.6 as the leading open-source model on several agentic and coding benchmarks, and competitive with closed frontier models from OpenAI, Anthropic, and Google DeepMind:[23]
| Benchmark | Kimi K2.6 | GPT-5.4 | Claude Opus 4.6 | Gemini 3.1 Pro | Kimi K2.5 |
|---|---|---|---|---|---|
| Humanity's Last Exam (full, with tools) | 54.0 | 52.1 | 53.0 | 51.4 | 50.2 |
| SWE-Bench Pro | 58.6 | 57.7 | 53.4 | 54.2 | 50.7 |
| SWE-Bench Verified | 80.2 | n/a | ~80.8 | n/a | 76.8 |
| Terminal-Bench 2.0 | 66.7 | 65.4 | 65.4 | 68.5 | n/a |
| LiveCodeBench v6 | 89.6 | n/a | 88.8 | n/a | ~85.0 |
| BrowseComp (Swarm mode) | 86.3 | n/a | n/a | n/a | 78.4 |
| DeepSearchQA (F1) | 92.5 | 78.6 | n/a | n/a | 77.1 |
| MathVision (with Python) | 93.2 | n/a | n/a | n/a | n/a |
K2.6 is available through kimi.com, the Kimi mobile app, the Moonshot API, and Kimi Code CLI. Weights were published on Hugging Face and GitHub under a Modified MIT License, the same permissive terms used for prior open-source K2 releases.[26] The model is also distributed through third-party inference platforms including Cloudflare Workers AI, vLLM, SGLang, and KTransformers.[31]
Kimi's user base grew rapidly following its launch, accelerated by the viral interest surrounding the 2-million-character context window upgrade in March 2024 and the Explore Edition launch in October 2024. At its peak around October 2024, Kimi reached over 36 million monthly active users (MAU), earning it the informal label of "China's ChatGPT" in some media reports.[17]
Moonshot AI invested heavily in user acquisition. Reports indicate that the company's monthly advertising spending peaked at nearly 200 million yuan (approximately $28 million), creating a "screen-dominating" trend on platforms like Bilibili, one of China's most popular video-sharing websites.[17] While this aggressive marketing successfully drove user numbers upward, it also drew scrutiny over the sustainability of such spending.[17]
Despite its strong early growth, Kimi faces intense competition in China's AI chatbot market. The primary competitive landscape as of early 2026 includes:
| Chatbot | Developer | Notable Strengths |
|---|---|---|
| Doubao | ByteDance | Largest user base in China (100M+ DAU by early 2026), integration with ByteDance ecosystem |
| DeepSeek | DeepSeek | Open-source models, strong reasoning, cost efficiency |
| Ernie Bot | Baidu | 200M+ MAU by January 2026, search engine integration, free access |
| Yuanbao | Tencent | Integration with WeChat ecosystem |
| Qwen | Alibaba | Strong open-source model family, cloud integration |
| Kimi | Moonshot AI | Long-context processing, open-source models, Agent Swarm |
The key differentiator is positioning: Doubao wins on raw distribution scale through ByteDance's Douyin and Toutiao apps, and DeepSeek competes on low-cost open-source reasoning, while Kimi has leaned into long-context processing, the open-weight K2 model family, and multi-agent orchestration via Agent Swarm. After peaking at 36 million MAU in October 2024, Kimi's user numbers fluctuated, dropping to roughly 10 to 15 million MAU by mid-2025 before recovering.[17] By October 2025, Kimi reached approximately 23.5 million global mobile MAUs, ranking among the top 20 AI apps worldwide by Sensor Tower data. In March 2026, kimi.com recorded 34.5 million website visits with a 47.9% bounce rate. ByteDance's Doubao, benefiting from its parent company's massive distribution network, became the clear market leader with nearly 60 million MAU by late 2024 and over 100 million DAU by early 2026. Kimi's positioning has shifted from raw scale (where it cannot match Doubao's organic distribution through Douyin and Toutiao) toward power users, developers, and the open-source community, where it has become one of the most cited Chinese frontier model families.
Moonshot AI began monetizing Kimi's API access in October 2025. The company employs multiple revenue streams: consumer subscription plans, API access for developers, and enterprise solutions.
The launch of Kimi K2.5 in January 2026 proved to be a financial inflection point. In fewer than 20 days following K2.5's release, Kimi's cumulative revenue exceeded its total revenue for the entire calendar year of 2025.[7] The revenue surge was driven by a spike in both global paying users and API usage.[6]
Overseas revenue has become a particularly important growth driver. Since November 2025, Kimi's overseas API revenue quadrupled, with monthly growth rates for both international and domestic paying users exceeding 170%.[18] By early 2026, overseas revenue had overtaken domestic income, with particularly strong international subscriber growth.[19] The April 2026 K2.6 release further accelerated API usage, with Cloudflare, OpenRouter, Fireworks, and several other inference providers making the open weights available within hours of the launch, broadening the addressable developer base beyond Moonshot's own platform.[31]
While Kimi was initially focused on the Chinese market, with primary support for Chinese language processing, the platform has progressively expanded its international footprint. The kimi.com web application is accessible globally, and the OK Computer agentic feature was first made available to international and Hong Kong users before mainland China.[12]
The open-source releases of Kimi K2 and K2.5 on platforms like Hugging Face and GitHub have been central to building an international developer community. Kimi K2's modified MIT license allows broad commercial and research use, attracting developers worldwide.[15]
Moonshot AI's strategic plans for 2026 and 2027 include expanding to international cloud regions, developing vertical agents for specialized domains such as law and medicine, building a community-driven tool marketplace, and advancing toward what the company calls "AGI Layer 2" with continual self-training architectures.
Kimi the consumer chatbot is a proprietary product, but the underlying Kimi K2 model family is released under open weights. Starting with Kimi K2 in July 2025, Moonshot AI has published full model weights on Hugging Face and GitHub under a Modified MIT License, the same permissive terms reused for K2 Thinking, K2.5, and K2.6.[15][26] The license permits broad commercial and research use, with an attribution requirement for products serving more than 100 million users or generating more than $20 million in monthly revenue.[15]
Moonshot AI's technical strategy has consistently emphasized a few core principles that differentiate it from competitors:
Long context as a foundation. Yang Zhilin has argued that long-context capabilities are not merely a feature but a fundamental requirement for useful AI systems, stating that "every problem is essentially a long-context problem."[33] Processing entire books, research papers, and codebases in a single pass enables applications that shorter-context models cannot support, such as comprehensive document analysis, cross-referencing across hundreds of pages, and maintaining coherent reasoning over very long chains of thought.
Open-source commitment. Starting with Kimi K2, Moonshot AI has released model weights openly, making its trillion-parameter models available for download and modification.[15] This approach, following the precedent set by companies like Meta with LLaMA and DeepSeek with its R1 model, has helped Moonshot build developer mindshare outside of China.
Agentic intelligence. The progression from K2 to K2 Thinking to K2.5 to K2.6 reflects a deliberate shift toward models that can act autonomously: using tools, calling APIs, browsing the web, writing and executing code, and coordinating large numbers of sub-agents. The Agent Swarm paradigm, introduced in K2.5 with up to 100 sub-agents and scaled to 300 in K2.6, represents Moonshot's most ambitious step in this direction.[23] Yang Zhilin has framed this with what he called the "three bets" of the K2 series at NVIDIA GTC 2026: training efficiency, long-context capability, and multi-agent coordination.
Efficient training. Kimi K2 Thinking's reported $4.6 million training cost drew attention for being significantly lower than the hundreds of millions of dollars reportedly spent on competing frontier models, although Moonshot itself stressed that the figure was not official.[14][34] Moonshot AI has invested heavily in training infrastructure optimization, including the development of custom optimizers like MuonClip.
| Date | Event |
|---|---|
| March 2023 | Moonshot AI founded by Yang Zhilin, Zhou Xinyu, and Wu Yuxin[1] |
| June 2023 | Seed funding (~$200M) from HongShan and ZhenFund at $300M valuation |
| October 2023 | First version of Kimi released (200,000 Chinese character context)[2] |
| November 16, 2023 | Kimi released to the general public[2] |
| February 2024 | Series B funding ($1B) led by Alibaba at $2.5B valuation[3][4] |
| March 2024 | Context window expanded to 2 million Chinese characters; server outage on March 21[16] |
| July 2024 | Context caching feature enters public beta |
| August 2024 | Series C funding ($300M) with Tencent at $3.3B valuation |
| October 11, 2024 | Kimi Explore Edition launched; MAU exceeds 36 million |
| January 20, 2025 | Kimi K1.5 released (reasoning model with RL training)[8] |
| June 2025 | Kimi-Researcher deep research agent released |
| July 2025 | Kimi K2 released (1T parameter open-source MoE model)[9] |
| September 9, 2025 | K2 updated (Kimi-K2-Instruct-0905); context expanded to 256K tokens |
| September 24, 2025 | OK Computer agentic feature announced[12] |
| October 2025 | API monetization begins |
| November 2025 | Kimi K2 Thinking released (open-source reasoning model)[14] |
| December 2025 | Series D funding ($500M) at $4.3B valuation[5] |
| January 27, 2026 | Kimi K2.5 released with Agent Swarm and multimodal vision[22] |
| February 2026 | Series E funding ($700M+) at $10B valuation; revenue in 20 days exceeds all of 2025[6][7] |
| March 2026 | Series E (Tranche 2) announced at $18B valuation; Hong Kong IPO discussions reported[27][28] |
| April 20, 2026 | Kimi K2.6 released and open-sourced; Agent Swarm scaled to 300 sub-agents and 4,000 steps; Skills feature introduced[23] |