Moonshot AI: Builder of Kimi K2, Valued at $20B (2026)

Moonshot AI is a Beijing AI lab founded in 2023 by Yang Zhilin, maker of the Kimi assistant and open-weight Kimi K2 coding models, valued at $20B in 2026.

Moonshot AI is a Beijing research lab founded in 2023 by ex-Google Brain researcher Yang Zhilin. It makes the Kimi consumer assistant and the Kimi K2 family of open-weight, trillion-parameter MoE models, the latest being Kimi K2.7 Code (June 2026), a 1T/32B-active coding model. Backed by Alibaba, Tencent, and a $2B Series D at a $20B valuation.

Moonshot AI, founded in March 2023 in Beijing by Yang Zhilin, builds the Kimi assistant and the open-weight Kimi K2 model family. The company raised a $2 billion Series D in May 2026 at a $20 billion valuation, backed by Alibaba and Tencent. Its latest model, Kimi K2.7 Code, is a 1-trillion-parameter MoE released under a Modified MIT license, priced at $0.95/$4.00 per million tokens.

Founded: 2023 · HQ: Beijing, China · Team: 200-300 · CEO: Yang Zhilin · Funding: $2B+ total raised, including a $2B Series D (May 7, 2026) led by Long-Z Investments (Meituan) · Valuation: $20B (Series D, May 2026); reportedly in talks for $30B as of June 2026

About Moonshot AI

Moonshot AI was founded in March 2023 in Beijing by Yang Zhilin together with Tsinghua University schoolmates Zhou Xinyu and Wu Yuxin. Yang Zhilin previously worked as an NLP researcher across Carnegie Mellon and Google Brain, where as first author of XLNet he helped build a pretraining model that beat Google's BERT on 20 of 20 benchmark tasks and earned an oral presentation slot at NeurIPS 2019. The company, known in Chinese as Yuezhi Anmian (Dark Side of the Moon), set out to build long-context, agentic large language models at a time when most Chinese labs were focused on shorter-context chat assistants. Within three months of founding, Yang had raised $60 million and assembled a roughly 40-person research team in Beijing, with a smaller secondary office later opened in Shanghai. The company's flagship consumer product is the Kimi assistant, a chat app that became one of the most downloaded AI apps in China on the strength of its long-context retrieval and document analysis. On the model side, Moonshot built the Kimi K2 family: a 1-trillion-parameter Mixture-of-Experts architecture with roughly 32 billion active parameters per token, a 256K-token context window, and Multi-head Latent Attention for efficient inference. K2 is released under a Modified MIT license that permits commercial use, including for very large deployments, which has made it one of the most widely adopted open-weight model families on Hugging Face. Moonshot has shipped a new major K2 release roughly every two months since mid-2025. Kimi K2 launched in July 2025 as the base agentic model. Kimi K2 Thinking followed in November 2025, adding visible reasoning and scoring 44.9% on Humanity's Last Exam and 60.2% on BrowseComp while running up to 200-300 sequential tool calls unattended. Kimi K2.5 arrived in January 2026 with native multimodal vision and an Agent Swarm mode coordinating up to 100 sub-agents. Kimi K2.6 followed in April 2026, scaling Agent Swarm to 300 domain-specialized sub-agents executing up to 4,000 coordinated steps in a single run. Kimi K2.7 Code launched June 12, 2026, a coding-specialist release on the same 1T/32B-active architecture that cuts thinking-token usage by roughly 30% versus K2.6 while reporting gains of 21.8% on Moonshot's internal Kimi Code Bench v2. Funding has scaled with the model cadence. Moonshot closed a $2 billion Series D on May 7, 2026, led by Long-Z Investments (the venture arm of food-delivery giant Meituan), with participation from state-backed China Mobile, Tsinghua Capital, and CPE Yuanfeng, alongside existing backers Alibaba, Tencent, HongShan, ZhenFund, IDG Capital, and 5Y Capital. That round valued the company at roughly $20 billion, up from about $4.3 billion at the end of 2025, making Moonshot the best-funded large language model startup in China. As of early June 2026, reports indicate Moonshot is in talks to raise a further $2 billion at a $30 billion valuation. Revenue comes from two channels: the Kimi consumer app (free with usage limits, plus paid tiers for higher quotas) and the Moonshot API platform, which bills per token for the K2 model family (Kimi K2.7 Code lists at $0.95 per million input tokens, $0.19 on a cache hit, and $4.00 per million output tokens). Because K2 weights are released under a permissive Modified MIT license, large enterprises can also self-host the models, which Moonshot positions as a complement to its hosted API rather than a competing revenue stream, similar to the open-weight strategy used by Meta and DeepSeek. Team size has grown from roughly 40 researchers at founding to an estimated 200-300 employees by 2026, concentrated in Beijing with a smaller Shanghai presence. Yang Zhilin remains CEO and the company's public face, frequently interviewed about Moonshot's research priorities. Compared to peers, Moonshot has stayed a relatively flat, research-heavy organization rather than building out a large enterprise sales function, which shows in its API-first go-to-market and reliance on third-party inference providers (Together, Fireworks, DeepInfra, OpenRouter, Azure, and others) to reach global developers. Moonshot's research output is published openly: technical reports for K2 and its successors appear on arXiv and Hugging Face, and the company actively promotes benchmark results through its own Kimi Code Bench, Program Bench, and MLS Bench suites. However, as of June 2026 Moonshot has not published a formal safety system card for any K2 release, has no disclosed third-party red-team partners, and has not announced a responsible scaling policy, in contrast to Anthropic, OpenAI, and Google DeepMind. Competitively, Moonshot sits alongside DeepSeek and Alibaba's Qwen team as one of China's three highest-profile open-weight model makers. Against DeepSeek, Moonshot has differentiated on agentic tool-use and long-horizon coding workflows (Agent Swarm) rather than raw reasoning benchmarks. Against closed frontier labs like Anthropic and OpenAI, Moonshot's K2 models trade a few points of benchmark performance for an order-of-magnitude lower API price and full self-hosting rights, though Moonshot has not yet published independently verified scores for K2.7 Code on standard suites like SWE-bench Verified or GPQA Diamond. On governance, Moonshot is a Beijing-incorporated private company with no disclosed EU entity or formal GDPR, SOC 2, or ISO 27001 certification. Its privacy policy states that user prompts and uploads may be used to improve and train its models, and independent security researchers (Wiz Research, NowSecure, SecurityScorecard) have reported issues including an exposed ClickHouse database and hardcoded encryption keys in Kimi's mobile app. For regulated or EU-based customers, third-party AI governance advisors recommend self-hosting the open-weight K2 models rather than relying on Moonshot's hosted API. With a new K2 release roughly every two months and a possible $30 billion valuation round in progress, Moonshot's near-term priority is clearly sustaining its release cadence and developer mindshare rather than enterprise compliance buildout.

Mission

Build long-context, agentic large language models and release them openly to push the boundary of what AI assistants can do for everyday knowledge work.

Products

Links

Website · GitHub · Twitter · LinkedIn · Blog · Docs

Frequently Asked Questions

What is Moonshot AI and what do they build?

Moonshot AI is a Beijing-based artificial intelligence lab founded in March 2023 by Yang Zhilin, a former NLP researcher at Carnegie Mellon and Google Brain who co-created the XLNet pretraining model. The company builds the Kimi consumer chat assistant and the Kimi K2 family of large language models, which use a Mixture-of-Experts architecture with roughly 1 trillion total parameters and 32 billion active per token. Its flagship 2026 release, Kimi K2.7 Code (June 12, 2026), is a coding-specialist model with a 256K-token context window released under a Modified MIT license. Moonshot reports gains of 21.8% on its own Kimi Code Bench v2 versus the prior K2.6 release, alongside roughly 30% fewer reasoning tokens per task. Users can access Kimi through its web and mobile app, or call the K2 models via the Moonshot API platform, OpenRouter, Together, Fireworks, and several other inference providers. Moonshot is positioned as one of China's leading open-weight model labs alongside DeepSeek and Alibaba's Qwen team.

Who founded Moonshot AI and who is the CEO?

Moonshot AI was founded in March 2023 by Yang Zhilin together with Tsinghua University schoolmates Zhou Xinyu and Wu Yuxin. Yang Zhilin remains the company's CEO and most public figure as of 2026. Before founding Moonshot, Yang was an NLP researcher who worked across Carnegie Mellon University and Google Brain, where as first author of the XLNet paper he helped build a model that beat Google's own BERT on all 20 benchmark tasks it was tested against and earned an oral presentation at NeurIPS 2019. Within three months of founding Moonshot, Yang raised $60 million in seed funding and assembled a roughly 40-person research team in Beijing. The company has since opened a smaller secondary office in Shanghai. No major leadership changes have been reported since founding, and Moonshot has not disclosed a separate president, CTO, or chief safety officer.

How much funding has Moonshot AI raised?

Moonshot AI's largest disclosed round is a $2 billion Series D closed on May 7, 2026, led by Long-Z Investments, the venture arm of Meituan, with participation from state-backed China Mobile, Tsinghua Capital, and CPE Yuanfeng. That round valued the company at roughly $20 billion, up sharply from about $4.3 billion at the end of 2025, making Moonshot the best-funded large language model startup in China. Earlier backers on the cap table include Alibaba, Tencent, HongShan (formerly Sequoia China), ZhenFund, IDG Capital, and 5Y Capital, alongside the original $60 million seed round Yang Zhilin raised within three months of founding in 2023. As of early June 2026, multiple reports indicate Moonshot is in talks to raise a further $2 billion at a $30 billion valuation. Moonshot has not disclosed revenue or ARR figures publicly. There is no announced IPO timeline or acquisition activity as of mid-2026.

What products does Moonshot AI make?

Moonshot AI's consumer-facing product is Kimi, a free chat assistant app (with paid usage tiers) known for long-context document analysis and agentic task execution. On the model side, Moonshot ships the Kimi K2 family on a roughly two-month cadence: Kimi K2 (July 2025), Kimi K2 Thinking (November 2025, with visible reasoning and up to 300 sequential tool calls), Kimi K2.5 (January 2026, native multimodal vision and 100-agent Agent Swarm mode), Kimi K2.6 (April 2026, Agent Swarm scaled to 300 sub-agents and 4,000 steps), and Kimi K2.7 Code (June 2026, a coding-focused release with roughly 30% fewer reasoning tokens). All K2 models are released under a Modified MIT license on Hugging Face for self-hosting, and are also available via the Moonshot API platform at $0.95 per million input tokens and $4.00 per million output tokens for K2.7 Code. The company also ships Kimi Code, a CLI-based coding agent built on K2.7 Code.

Where is Moonshot AI headquartered and how big is the team?

Moonshot AI is headquartered in Beijing, China, with a smaller secondary office in Shanghai. The team has grown from roughly 40 researchers assembled within three months of the company's March 2023 founding to an estimated 200-300 employees by 2026. The company has stayed relatively lean and research-heavy compared to peers, without a large publicly visible enterprise sales organization. Most of Moonshot's engineering and research headcount is concentrated in Beijing, where the founding team (Yang Zhilin, Zhou Xinyu, and Wu Yuxin, all Tsinghua University alumni) is based. Moonshot has not disclosed a remote-work policy, and public job listings are concentrated in mainland China via its careers.kimi.com site. There have been no reported layoffs or major restructuring events as of June 2026; headcount growth has tracked the company's funding rounds and accelerating release cadence.

What is Moonshot AI's mission or research focus?

Moonshot AI's stated focus is building long-context, agentic large language models and releasing them openly so developers can use and self-host them. In practice this shows up as the Kimi K2 family's 256K-token context window, Mixture-of-Experts architecture (1 trillion total parameters, 32 billion active per token), and a string of Agent Swarm releases that let a model coordinate up to 300 sub-agents across thousands of coordinated steps. Moonshot publishes technical reports on arXiv and Hugging Face for each major K2 release and runs its own benchmark suites, including Kimi Code Bench v2, Program Bench, and MLS Bench Lite, which it used to report 21.8%, 11.0%, and 31.5% gains respectively for K2.7 Code over K2.6. Unlike Anthropic's Constitutional AI or OpenAI's Preparedness Framework, Moonshot has not published a formal responsible scaling policy, safety system card, or named third-party red-team partners for any K2 release as of mid-2026.

Is Moonshot AI compliant with SOC 2, GDPR, HIPAA?

No. As of mid-2026, Moonshot AI has not published any SOC 2, ISO 27001, or HIPAA certification, and has no disclosed trust center. Moonshot's own privacy policy for the Kimi API platform states that user prompts and uploaded content may be used to improve and train its models, and that personal information may be shared with service providers and affiliates, with no documented opt-out or zero-retention enterprise tier. Data processed through Moonshot's hosted API and Kimi app is handled in China, which AI governance advisors flag as a blocker for GDPR compliance in the EU unless a customer self-hosts the open-weight K2 models on their own infrastructure. Independent security researchers, including Wiz Research, NowSecure, and SecurityScorecard, have separately reported a publicly exposed ClickHouse database containing chat logs and API keys, unencrypted device data transmission in the iOS app, and hardcoded encryption keys. Moonshot has no disclosed EU AI Act classification.

Who are Moonshot AI's main competitors?

Moonshot AI's closest competitors are other Chinese open-weight labs, primarily DeepSeek and Alibaba's Qwen team. Against DeepSeek, Moonshot differentiates on agentic tool-use and long-horizon coding workflows through its Agent Swarm system, which scales to 300 sub-agents and 4,000 coordinated steps, while DeepSeek has generally published stronger independently verified reasoning benchmark scores. Against Alibaba's Qwen models, Moonshot competes directly on open-weight agentic coding, with both labs releasing models under permissive licenses for self-hosting. Against closed frontier labs like Anthropic and OpenAI, Moonshot's K2 models trade some benchmark performance, particularly on independently verified suites like SWE-bench Verified and GPQA Diamond where K2.7 Code has no published third-party scores, for an order-of-magnitude lower API price ($0.95/$4.00 per million tokens) and full self-hosting rights. A category where Moonshot has no real answer yet is enterprise trust and compliance, where it lags Anthropic and OpenAI on certifications and safety documentation.