CtrlK
| Moonshot AI | |
|---|---|
| 北京月之暗面科技有限公司 | |
| * | |
| File:Moonshot AI Headquarters.jpg | |
| Moonshot AI headquarters in Beijing | |
| Type | Private company |
| Industry | Artificial intelligence |
| Founded | March 2023 |
| Founders | Yang Zhilin (CEO) Zhou Xinyu Wu Yuxin |
| Headquarters | 13th Floor, Building 1, JD Technology Building No. 76 Zhichun Road, Haidian District, Beijing, China |
| Key people | Yang Zhilin (CEO) Zhou Xinyu (Co-founder) Wu Yuxin (Co-founder) Zhang Yutao (Product Development) |
| Products | Kimi AI assistant Kimi K2.5 model Kimi K2 model Kimi K1.5 model Mooncake platform Kimi Slides Moonlight models Ohai (social AI app) Noisee (audio generation tool) |
| Valuation | $18 billion (March 2026) |
| Employees | ~300 (2025) |
| Website | moonshot.cn |
Moonshot AI (Chinese: 月之暗面; pinyin: Yuè Zhī Ànmiàn, literally "Dark Side of the Moon"; legal name: Beijing Moonshot AI Technology Co., Ltd.) is a Chinese artificial intelligence company headquartered in Beijing, founded in March 2023. The company specializes in developing large language models (LLMs) and aims to achieve artificial general intelligence (AGI) through foundational models with capabilities in long context processing, multimodal world modeling, and scalable architectures for self-improvement.[1] Moonshot AI has been dubbed one of China's "AI Tigers" or as part of the "Six Tigers" by investors, alongside companies like Zhipu AI, MiniMax, 01.AI, Baichuan AI, StepFun, and DeepSeek.[2] The company is best known for its Kimi chatbot and its series of open-source models, including Kimi K2, Kimi K2.5, and the reasoning-focused Kimi K1.5, all of which feature advanced long-context processing capabilities.
As of March 2026, Moonshot AI has raised over $2.5 billion in total funding, with its latest reported valuation reaching approximately $18 billion, making it one of the fastest-growing AI startups in the world.[3] The company became the fastest Chinese startup to achieve decacorn status (a valuation exceeding $10 billion), reaching that milestone in under three years from founding.[4]
Moonshot AI was founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, all alumni of Tsinghua University, amid a surge in interest in generative AI following the success of ChatGPT.[5] The company was launched on the 50th anniversary of Pink Floyd's album The Dark Side of the Moon, which was founder Yang Zhilin's favorite album and the inspiration for the company's Chinese name.[1] The English name "Moonshot" reflects the ambitious and difficult nature of the company's mission, described by Yang as "like landing on the moon."[6]
Yang Zhilin stated that his goal for founding Moonshot AI was to build foundational models to achieve artificial general intelligence (AGI), with three key milestones: long context length, multimodal world model, and a scalable general architecture capable of continuous self-improvement without human input.[1]
In October 2023, Moonshot launched its first product, the Kimi chatbot, to the public. The chatbot initially could process up to 200,000 Chinese characters per conversation, making it the world's first AI assistant with such extensive context-handling capabilities.[5]
By March 2024, Moonshot claimed Kimi could handle 2 million Chinese characters in a single prompt, a significant upgrade from the previous version.[7] The chatbot emerged as the closest rival to Baidu's Ernie Bot in the Chinese market. Due to increased user demand following this upgrade, Kimi suffered a two-day outage on March 21, 2024, prompting a public apology from the company.[1]
By the end of 2024, Kimi had reached more than 36 million monthly active users across web, app, and mini-programs, establishing it as the second-most-popular AI chatbot in China behind ByteDance's Doubao.[8]
On January 20, 2025, Kimi K1.5 was released, which Moonshot claimed matched the performance of OpenAI's o1 model in mathematics, coding, and multimodal reasoning capabilities. The accompanying technical paper described a simplified reinforcement learning framework that achieved state-of-the-art results without relying on Monte Carlo tree search or process reward models.[9]
In April 2025, Moonshot released Kimi-VL, an open-source 16 billion parameter mixture-of-experts vision-language model with 3 billion active parameters, designed for multimodal reasoning, long-context understanding, and agent capabilities. A reasoning variant, Kimi-VL-Thinking, followed in June 2025.[10]
In June 2025, Moonshot AI released Kimi-Dev-72B, a 72-billion-parameter model specifically designed for software engineering tasks. Built on the Qwen 2.5-72B foundation and optimized through reinforcement learning, the model achieved a 60.4% resolve rate on SWE-bench Verified, setting a new state-of-the-art among open-source models at the time of release.[11]
In July 2025, Moonshot AI released the weights for Kimi K2, a large language model with 1 trillion total parameters using a mixture-of-experts (MoE) architecture, trained on 15.5 trillion tokens and released under a modified MIT License.[12]
On September 9, 2025, Moonshot AI released Kimi-K2-Instruct-0905, which increased performance in agentic coding tasks and doubled the context window to 256K tokens.[13]
On November 6, 2025, Moonshot AI launched Kimi K2 Thinking, a reasoning-focused update to K2 that represented the first generation thinking agent with native support for "thinking while using tools." The model could execute 200 to 300 sequential tool calls without human interference, reasoning coherently across hundreds of steps to solve complex problems. It scored 43% on Humanity's Last Exam, reportedly exceeding the performance of OpenAI's GPT-5 and Anthropic's Claude Sonnet 4.5 on that benchmark. The model was reported to have cost only $4.6 million to train.[14]
In December 2025, Moonshot AI closed a $500 million Series C round led by IDG Capital, with participation from Alibaba Group and Tencent Holdings, bringing the company's valuation to $4.3 billion and its cash reserves to over 10 billion yuan (approximately $1.4 billion).[15]
On January 27, 2026, Moonshot AI released Kimi K2.5, its most powerful model to date, featuring native multimodal capabilities through a 400-million-parameter vision encoder called MoonViT. The model introduced Agent Swarm technology, allowing it to coordinate up to 100 specialized AI agents working simultaneously. K2.5 scored 50.2% on Humanity's Last Exam (HLE) with tools and 78.4% on BrowseComp in Agent Swarm mode.[16]
Following the release of K2.5, the company experienced explosive revenue growth. In fewer than 20 days after K2.5's launch, Kimi's cumulative revenue surpassed the company's total revenue for all of 2025, with overseas revenue overtaking domestic income for the first time.[4] Monthly growth rates for both overseas and domestic paying users exceeded 170%.[17]
In February 2026, Moonshot secured over $700 million in new funding co-led by Alibaba Group, Tencent Holdings, Wuyuan Capital, and Ji'an Investment, bringing the company's valuation to approximately $10 billion. This made Moonshot AI the fastest Chinese startup to achieve decacorn status.[4][17]
By March 2026, reports indicated that Moonshot AI was seeking to raise an additional $1 billion at a valuation of approximately $18 billion, more than quadrupling its valuation in the span of just three months.[3]
Yang Zhilin (杨植麟) serves as CEO and co-founder. He holds a Ph.D. from Carnegie Mellon University's Language Technologies Institute, where he completed the program in under four years, studying under Ruslan Salakhutdinov (Apple's director of AI research) and William Cohen (Google DeepMind principal scientist).[5] He graduated from Tsinghua University, where he ranked first in his class at the Computer Science Department.[5] During his academic career, Yang collaborated with multiple Turing Award winners, publishing over 20 influential papers, including co-authored work with Yoshua Bengio and Yann LeCun.[18]
Before founding Moonshot AI, Yang worked at Facebook AI Research (now Meta AI) and Google Brain, where he co-authored influential papers including Transformer-XL and XLNet.[5] He also worked with Huawei Technologies on an early version of the Pangu AI model in 2020 and led a team to develop the Wudao LLM at the Beijing Academy of Artificial Intelligence in 2021.[5]
Yang has described his vision for Moonshot AI as combining "the technology idealism of OpenAI and business philosophy of ByteDance."[5] He has also stated: "We don't want to be anything Chinese, nor necessarily OpenAI," arguing that a truly impactful AGI company cannot endure long-term if confined to a regional market.[19]
Zhou Xinyu (周昕宇) is a co-founder who previously worked at Hulu, Tencent, and Megvii, conducting research in deploying deep neural networks on hardware with limited computational resources.[19]
Wu Yuxin (吴育昕) is a co-founder who previously worked at Google Brain on foundation models and at Meta AI Research on computer vision.[19]
Kimi is Moonshot AI's flagship consumer product, an AI assistant chatbot launched in October 2023. The name comes from Yang Zhilin's English nickname.[1] The chatbot's defining feature has been its industry-leading context window, which has expanded significantly over time:
| Version | Date | Context Window | Notes |
|---|---|---|---|
| Kimi (initial) | October 2023 | 200,000 Chinese characters | First AI chatbot with this context length |
| Kimi (upgraded) | March 2024 | 2 million Chinese characters | Equivalent to several novels in a single prompt |
| Kimi K1.5 | January 2025 | 128K tokens | Reasoning model with long and short CoT modes |
| Kimi K2 | July 2025 | 128K tokens (256K in K2-Instruct-0905) | MoE architecture, 1T parameters |
| Kimi K2 Thinking | November 2025 | 256K tokens | Native "thinking while using tools" support |
| Kimi K2.5 | January 2026 | 128K tokens | Native multimodal, Agent Swarm |
This long-context capability allows the chatbot to analyze and summarize the content of lengthy documents, such as academic papers, financial reports, or entire books, in a single query. For comparison, 2 million Chinese characters is roughly equivalent to the text of several novels.[7]
In China, Kimi offers six tiers of subscription plans ranging from 5.2 yuan for four days to 399 yuan for a year of priority access.[20]
By the end of 2024, Kimi had reached over 36 million monthly active users. The user base experienced fluctuations throughout 2025, dropping to approximately 9.9 million monthly active users in Q3 2025 before recovering to 23.6 million in December 2025.[8] Despite these fluctuations in total user count, the company saw strong growth in paying subscribers, particularly following the release of K2.5 in January 2026.
Moonshot AI has released a series of models with increasing capabilities. The following table summarizes the major model releases:
| Model | Release Date | Parameters | Architecture | Key Features | License |
|---|---|---|---|---|---|
| Kimi K1.5 | January 2025 | ~500B (estimated) | Dense Transformer | Reinforcement learning-trained reasoning; long-CoT and short-CoT modes; 128K context; multimodal (text + vision) | Proprietary |
| Kimi-VL (A3B) | April 2025 | 16B total / 3B active | MoE Vision-Language | Multimodal reasoning, long-context understanding, agent capabilities | Open-source |
| Kimi-Dev-72B | June 2025 | 72B | Dense Transformer (Qwen 2.5-72B base) | Software engineering tasks; 60.4% on SWE-bench Verified; BugFixer + TestWriter framework | Open-weight |
| Kimi K2 | July 2025 | 1T total / 32B active | MoE (384 experts, 8 selected) | Agentic coding, 128K context, 15.5T training tokens | Modified MIT |
| Kimi-K2-Instruct-0905 | September 2025 | 1T total / 32B active | MoE | Improved coding performance, 256K context | Modified MIT |
| Kimi K2 Thinking | November 2025 | 1T total / 32B active | MoE | Native tool-use during reasoning; 200-300 sequential tool calls; INT4 quantization | Open-weight |
| Moonlight (3B/16B) | 2025 | 3B / 16B | MoE | Muon optimizer, 5.7T training tokens | Open-source |
| Kimi K2.5 | January 2026 | 1T total / 32B active | MoE (384 experts) | Native multimodal via MoonViT (400M params); Agent Swarm (up to 100 agents); Instant/Thinking/Agent/Agent Swarm modes | Open-weight |
Kimi K1.5 was released on January 20, 2025, as Moonshot AI's first dedicated reasoning model. The model was trained using a simplified reinforcement learning framework that eschewed complex techniques such as Monte Carlo tree search, value functions, and process reward models in favor of a streamlined approach based on scaling RL with LLMs.[9]
The model operates in two modes: a long chain-of-thought (long-CoT) mode optimized for detailed step-by-step reasoning, and a short chain-of-thought (short-CoT) mode for concise answers. Key benchmark results include:
| Benchmark | Long-CoT | Short-CoT | Comparison |
|---|---|---|---|
| AIME | 77.5 | 60.8 | Matches OpenAI o1 (long-CoT) |
| MATH 500 | 96.2 | 94.6 | Outperforms GPT-4o by large margin (short-CoT) |
| Codeforces | 94th percentile | N/A | Competitive with top reasoning models |
| MathVista | 74.9 | N/A | Exceeds OpenAI o1's 63.8 on vision tasks |
| LiveCodeBench | N/A | 47.3 | Outperforms Claude Sonnet 3.5, GPT-4o |
The accompanying technical paper, "Kimi k1.5: Scaling Reinforcement Learning with LLMs" (arXiv:2501.12599), demonstrated that scaling the context window of RL to 128K tokens led to continued performance improvements, and that joint training on text and vision data produced a model capable of reasoning across both modalities.[9]
In July 2025, Moonshot AI released the weights for Kimi K2, a large language model with 1 trillion total parameters using a mixture-of-experts (MoE) architecture, where 32 billion parameters are active during inference.[21] The model was trained on 15.5 trillion tokens of data and is released under a modified MIT License.[1]
Key features of Kimi K2 include:
| Feature | Specification |
|---|---|
| Total Parameters | 1 trillion |
| Active Parameters | 32 billion |
| Architecture | Mixture-of-Experts (MoE) with 384 experts |
| Training Data | 15.5 trillion tokens |
| Context Window | 128K tokens (256K in K2-Instruct-0905) |
| License | Modified MIT License |
The model achieved state-of-the-art performance among open-source non-thinking models, with notable scores including:[22]
Released on November 6, 2025, Kimi K2 Thinking was the first generation thinking agent with native support for reasoning while using tools. Built on the same 1-trillion-parameter MoE architecture as K2, the model used native INT4 quantization with a 256K context window, achieving lossless reductions in inference latency and GPU memory usage.[14]
The model demonstrated several notable capabilities:
Model weights were published on Hugging Face (moonshotai/Kimi-K2-Thinking) to support local deployment.
Released on January 27, 2026, Kimi K2.5 is Moonshot AI's most capable model as of early 2026. It was built through continual pretraining on approximately 15 trillion mixed visual and text tokens on top of the Kimi K2 base model. The major addition is native multimodal support via MoonViT, a 400-million-parameter vision encoder that processes images through the same transformer architecture as text, rather than relying on a grafted adapter.[16]
Kimi K2.5 supports four operational modes:
| Mode | Description |
|---|---|
| Instant | Fast responses for straightforward queries |
| Thinking | Extended reasoning with chain-of-thought for complex problems |
| Agent | Single-agent tool use for multi-step tasks |
| Agent Swarm | Coordination of up to 100 specialized AI agents working in parallel |
The Agent Swarm feature represents a distinctive innovation, enabling the model to parallelize complex tasks across many specialized sub-agents. In practice, this parallel approach reduces execution time by 4.5 times compared to single-agent processing while achieving strong benchmark results at 76% lower cost than comparable frontier models.[16]
Key benchmark results for Kimi K2.5:
| Benchmark | Score | Mode | Comparison |
|---|---|---|---|
| Humanity's Last Exam (HLE) | 50.2% | Agent (with tools) | 18.2 points above Claude Opus 4.5's 32.0% |
| BrowseComp | 78.4% | Agent Swarm | Above Claude Opus 4.5's 65.8% |
| BrowseComp | 74.9% | Standard Agent | Single agent also outperforms competitors |
| DeepSearchQA | 77.1% | Agent | Multi-step information retrieval |
| WideSearch F1 | 79.0% | Agent Swarm | Up from 72.7% in single-agent mode |
Model weights and code are publicly available on Hugging Face and GitHub under an open-weight license.[16]
Mooncake is the serving platform for Moonshot's Kimi chatbot, processing over 100 billion tokens daily.[1] It features a KVCache-centric disaggregated architecture that separates prefill and decoding clusters, utilizing underused CPU, DRAM, SSD, and NIC resources of GPU clusters.[23]
Moonshot was awarded the Erik Riedel Best Paper Award at the USENIX FAST conference in February 2025 for the paper detailing Mooncake's architecture.[24] In practical deployments, Mooncake enabled Kimi to handle 115% and 107% more requests on NVIDIA A800 and H800 clusters, respectively, compared to previous systems.[23]
Key components of Mooncake include:
The platform's code has been partially open-sourced on GitHub.[25]
In 2025, Moonshot AI launched Kimi Slides, an AI agent within the Kimi ecosystem that generates professional presentations from text prompts, documents, or website URLs.[26] The tool is part of Kimi+, Moonshot AI's premium version of the chatbot, though it was initially offered for free to all users.[26]
In late 2025, Moonshot AI introduced "OK Computer," an agent mode for Kimi that can create multi-page websites and editable slides from simple prompts. The feature supports processing up to 1 million rows of input data and multimedia outputs including text, audio, and video.[27] The name "OK Computer" is a reference to the 1997 album by Radiohead.[28]
Moonshot AI provides an open platform for developers via its API service at platform.moonshot.ai. This includes access to models like Kimi K1.5 and K2.5, which feature a 128K context window, function calling, and vision support. Pricing starts at $0.004 per 1K input tokens and $0.006 per 1K output tokens, with a free tier of 1 million tokens monthly.[29]
Moonshot AI has raised significant capital since its inception, establishing it as one of the most valuable AI startups in the world. The company's valuation grew from $300 million at its seed round in 2023 to approximately $18 billion by March 2026, an increase of roughly 60 times in under three years.
| Date | Round | Amount Raised | Lead Investors | Post-Money Valuation | Ref |
|---|---|---|---|---|---|
| 2023 | Seed | $60 million | N/A | $300 million | [1] |
| 2023 | Series A | > $300 million | HongShan (formerly Sequoia Capital China), Zhen Fund | Not specified | [30] |
| February 2024 | Series B | $1 billion | Alibaba Group, HongShan, Monolith Management, Xiaomi, Tom Preston-Werner | $2.5 billion | [31][32] |
| August 2024 | Series B Extension | $300 million | Tencent, Gaorong Capital | $3.3 billion | [33] |
| December 2025 | Series C | $500 million | IDG Capital, Alibaba, Tencent | $4.3 billion | [15] |
| February 2026 | Series D | > $700 million | Alibaba, Tencent, Wuyuan Capital, Ji'an Investment | ~$10 billion | [17] |
| March 2026 (reported) | Series E (planned) | ~$1 billion (target) | Not disclosed | ~$18 billion (target) | [3] |
The February 2024 round was the largest single funding round for Chinese LLM developers on public record at the time.[31] By March 2026, Moonshot AI's total cumulative funding exceeded $2.8 billion across all rounds.
Moonshot AI generated approximately $240 million in revenue through November 2025.[34] The company's revenue trajectory accelerated dramatically following the release of Kimi K2.5 in January 2026. In fewer than 20 days after the model's launch, Kimi's cumulative revenue already exceeded its total revenue for all of 2025.[4] A significant shift occurred as overseas revenue surpassed domestic income for the first time, with Kimi's overseas API revenue quadrupling since November 2025.[17]
In collaboration with UCLA, Moonshot AI researchers published "Muon is Scalable for LLM Training," demonstrating successful scaling of the Muon optimizer.[35] The Muon optimizer, which uses momentum orthogonalized by Newton-Schulz iterations, achieves approximately 2 times computational efficiency compared to AdamW under compute-optimal conditions.[36]
Key innovations include:
Based on the Muon optimizer research, Moonshot AI released Moonlight, a series of MoE models in 3B and 16B parameter configurations, trained with 5.7 trillion tokens.[37] The models demonstrated superior performance compared to similar-scale models while requiring significantly fewer training FLOPs.[36]
The Kimi k1.5 technical paper (arXiv:2501.12599) introduced several contributions to the field of reinforcement learning for language models. It demonstrated that a simplified RL framework, without reliance on Monte Carlo tree search, value functions, or process reward models, could achieve state-of-the-art reasoning performance. The paper also showed that scaling the context window of RL training to 128K tokens produced continued improvements, and that joint training on text and vision data yielded effective cross-modal reasoning.[9]
Moonshot AI is considered one of China's "Six Tigers" (also called "AI Tigers" or "Six Little Tigers"), a group of leading AI startups that emerged in 2023-2024 to challenge both domestic incumbents and international competitors. The Six Tigers are:[2]
| Company | Founded | Headquarters | Notable Product | Key Focus |
|---|---|---|---|---|
| Moonshot AI | 2023 | Beijing | Kimi chatbot | Long-context LLMs, multimodal AI |
| Zhipu AI | 2019 | Beijing | ChatGLM | Foundation models, enterprise AI |
| MiniMax | 2021 | Shanghai | Talkie (companion chatbot) | Consumer AI, global expansion |
| 01.AI | 2023 | Beijing | Yi models | Open-source LLMs |
| Baichuan AI | 2023 | Beijing | Baichuan models | Enterprise and consumer LLMs |
| StepFun | 2023 | Shanghai | Step models | Multimodal AI systems |
DeepSeek is sometimes included in an expanded version of this group, though it operates as a research lab funded by the quantitative hedge fund High-Flyer rather than as a traditional venture-backed startup.[2]
Moonshot AI competes with both domestic and international AI companies. In the Chinese market, its primary competitors include:
Internationally, Moonshot AI's models compete with offerings from OpenAI, Anthropic (Claude), Google (Gemini), and Meta (LLaMA).
Moonshot AI differentiated itself early by focusing on long-context processing, a capability that took many competitors several months to match. As of 2026, the company has further distinguished itself through its Agent Swarm technology and rapid open-source model releases.
Moonshot AI is headquartered at:
The company also maintains an office in Shanghai.[39]
As of 2025, Moonshot AI employs approximately 300 people, having grown from just 40 employees at the time of its initial funding in October 2023 and approximately 80 employees in February 2024.[39][40]
During China's June 2025 gaokao period, Moonshot AI, along with several major tech platforms including Baidu, ByteDance, and Tencent, temporarily restricted certain AI features to mitigate exam-related misuse and prevent cheating.[41]
In June 2024, reports suggested Moonshot AI was planning to expand into the US market with products like Ohai (a social AI app) and Noisee (an audio generation tool), but the company denied these intentions at the time.[42] By early 2026, however, international expansion had become a central part of the company's strategy. Overseas revenue surpassed domestic income following the release of K2.5, with particularly strong international subscriber growth. The company's API saw quadrupled overseas revenue between November 2025 and February 2026.[17]
Following the February 2024 funding round, reports emerged that CEO Yang Zhilin and related individuals cashed out $40 million in shares, an unusually large amount for a first-year startup, raising concerns among investors.[43]
In November 2024, a group of investors filed for arbitration against the company's co-founder and Chief Technology Officer, alleging that funding rounds were conducted without obtaining required consent from some AI-focused investors.[1] The dispute involved GSR Ventures China and four other firms that had invested in Yang Zhilin's previous venture, Recurrent AI.[44]
This led to reduced involvement from some investors and ongoing legal proceedings into 2025. Additional disputes involved alleged conflicts of interest related to a spin-off company and fiduciary breaches, further complicating investor relations.[45] As of February 2025, the arbitration case advanced without settlement.[46]
Moonshot AI has made several significant open-source contributions to the AI community:
| Project | Description | License | Repository |
|---|---|---|---|
| Kimi K2 | 1T-parameter MoE LLM | Modified MIT | Hugging Face, GitHub |
| Kimi K2 Thinking | Reasoning model with tool-use | Open-weight | Hugging Face |
| Kimi K2.5 | Native multimodal agentic model | Open-weight | Hugging Face, GitHub |
| Kimi-VL (A3B) | MoE vision-language model | Open-source | GitHub |
| Kimi-Dev-72B | Software engineering model | Open-weight | Hugging Face, GitHub |
| Moonlight (3B/16B) | MoE models trained with Muon optimizer | Open-source | GitHub |
| Mooncake | LLM serving platform (Transfer Engine) | Open-source | GitHub |
| Muon Optimizer | Efficient optimizer for LLM training | Open-source | GitHub |
Yang Zhilin has described his vision for Moonshot AI as combining "the technology idealism of OpenAI and business philosophy of ByteDance."[5] This philosophy aims to balance AGI's far-reaching potential with the need for practical, user-centered solutions that can sustain a commercially viable enterprise.[19]
The company's mission is inherently global in scope. Yang has stated: "We don't want to be anything Chinese, nor necessarily OpenAI," arguing that a truly impactful AGI company cannot endure long-term if confined to a regional market.[19]
