Moonshot AI

Updated June 2026

Moonshot AI is the Beijing lab behind Kimi, one of China's 'Six AI Tigers'. Founded in 2023 by Tsinghua and Carnegie Mellon researcher Yang Zhilin, it pioneered long-context chat and, in 2026, became the leading open-weight AI lab with its Kimi K2 models — raising about $2 billion at a $20 billion valuation while preparing for a Hong Kong IPO.

Website →API →Pricing →

Moonshot AI is the Beijing lab behind Kimi, the AI assistant and series of open-weight models that made it one of China’s “Six AI Tigers”. Founded in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin, it first made its name on long context — its 2023 chatbot was the first to handle 128,000 tokens — and in 2026 became the world’s leading open-weight AI lab. Its flagship, Kimi K2.6, is the strongest open-weights model on independent testing (Artificial Analysis, April 2026), released under a permissive modified MIT licence that lets anyone download and self-host it.

The 2026 story is a comeback built on open weights. After Kimi’s consumer popularity fell through 2025, Moonshot pivoted to open-sourcing frontier-adjacent models — a strategy that rebuilt its standing among developers, drew a fresh wave of capital, and made Kimi K2 “another DeepSeek moment” on release. In May 2026 the company raised about $2 billion at a roughly $20 billion valuation, led by Meituan’s Long-Z Investments, and is now restructuring to pursue a Hong Kong listing.

Quick facts

History and founding

Moonshot AI was incorporated in March 2023 in Beijing by Yang Zhilin, Zhou Xinyu and Wu Yuxin, schoolmates from Tsinghua University. The company is named after Pink Floyd’s The Dark Side of the Moon (月之暗面), Yang’s favourite album, and was launched on the record’s 50th anniversary. Yang has framed the company’s purpose as building foundation models toward artificial general intelligence, with three milestones: long context length, a multimodal world model, and a scalable architecture capable of continuous self-improvement.

Yang Zhilin is the technical face of the company. He took a computer-science degree at Tsinghua, then a PhD at Carnegie Mellon University under Ruslan Salakhutdinov and William Cohen, finishing in four years. He is a co-author of two influential NLP papers — Transformer-XL and XLNet, the latter of which outperformed Google’s BERT on 20 standard tasks and was an oral presentation at NeurIPS 2019 — and worked at Google Brain and Meta during his studies. That long-context research lineage is visible in Kimi’s founding bet.

Moonshot released the first version of Kimi in October 2023, capable of processing up to 200,000 Chinese characters, and opened it to the public on 16 November 2023 — the first assistant to support a 128,000-token context. A March 2024 beta pushed that to two million Chinese characters in a single prompt, the feature that first made Kimi a household name in China and triggered a two-day outage from demand.

Funding and valuation

Moonshot is backed by China’s largest internet companies and, increasingly, by state-linked capital. Its raises rank among the biggest in China’s large-language-model sector.

The May 2026 round at roughly $20 billion was the largest single financing event in China’s LLM sector at the time, and it marked a roughly fivefold jump in valuation in under six months — from about $4.3 billion at the end of 2025. Alibaba is reported to be Moonshot’s largest outside shareholder, and Tencent an early backer; the May 2026 round notably brought in state-backed China Mobile and Tsinghua Capital. Moonshot is reportedly seeking a further round of around $2 billion that would value it near $30 billion, though that figure is unconfirmed.

As a private Chinese company, Moonshot does not publish audited financials, so valuation and funding figures come from press reports rather than filings and should be read accordingly.

Models and the open-weight strategy

Moonshot’s model line is built around Mixture-of-Experts (MoE) architectures and, since mid-2025, an open-weight release strategy. The flagship is Kimi K2.6; the full lineup and live benchmarks render below this page.

The strategy turned on Kimi K2 in July 2025. Releasing a one-trillion-parameter model (32 billion active) under a modified MIT licence, Moonshot reclaimed developer attention it had lost as Kimi’s consumer numbers slid, and the model topped Hugging Face downloads on launch. The open-weight approach also fits a broader Chinese pattern of open-sourcing models to counter US efforts to limit China’s access to advanced chips. By April 2026, Kimi K2.6 was ranked the strongest open-weight model in the world by Artificial Analysis, within three points of the top US frontier models. See the best AI models for where it sits in the wider ranking.

A practical detail matters commercially: the modified MIT licence requires attribution only for products exceeding 100 million monthly active users or $20 million in monthly revenue, leaving the models free to download, fine-tune and self-host for almost everyone else.

Long context, Mooncake and Muon

Moonshot’s research signature is infrastructure for long context and efficient training. Mooncake, the platform that serves Kimi, processes around 100 billion tokens a day and won the Erik Riedel Best Paper Award at the USENIX FAST conference for its KVCache-centred, disaggregated architecture — work that helps Chinese labs do more with constrained hardware. In a joint paper with UCLA, Moonshot researchers showed the Muon optimiser could be scaled to train a 16-billion-parameter MoE model with roughly twice the computational efficiency of the standard AdamW optimiser, and open-sourced the implementation. This focus on squeezing more from each chip is a recurring theme for a lab operating under US export controls.

Products and ecosystem

• Kimi app — the consumer assistant (web, iOS, Android, desktop PWA), free to use with paid Moderato, Allegretto, Allegro and Vivace tiers ($19–$199/month). Its OK Computer agent mode builds websites, slides and data analyses.
• Moonshot Open Platform — the pay-as-you-go developer API at platform.moonshot.ai, where Kimi K2.6 costs roughly $0.95 per million input tokens and $4 per million output tokens.
• Kimi Code — an agentic coding command-line tool that defaults to the latest Kimi model, now backed by the June 2026 Kimi K2.7 Code model.
• Kimi Work — a local desktop agent for macOS and Windows, launched for testing in June 2026 and running on Kimi K2.6, pairing an Agent Swarm of up to 300 sub-agents with browser automation.
• Open weights — the Kimi K2 family on Hugging Face, the route most organisations use to run Kimi outside China’s jurisdiction.

Business and financials

Moonshot earns money three ways: the consumer Kimi subscriptions, the pay-as-you-go API on the Open Platform, and agentic coding through Kimi Code. The open weights are not a direct revenue line, but they drive adoption that feeds the paid API and the brand.

Revenue figures are not audited, but the reported trajectory is steep: Moonshot has said revenue from Kimi K2.5 in its first 20 days exceeded its entire 2025 total, and press reports in mid-2026 described revenue roughly doubling within six weeks on the back of K2.6 and Kimi Code. Treat these as company and press figures, not filings. The commercial logic is to convert open-weight credibility and a low-cost API into developer and enterprise spend, much as DeepSeek does.

Leadership

• Yang Zhilin — CEO and co-founder; Tsinghua and Carnegie Mellon researcher, co-author of Transformer-XL and XLNet, formerly at Google Brain and Meta.
• Zhou Xinyu — co-founder; previously a senior researcher at the computer-vision company Megvii (Face++).
• Wu Yuxin — co-founder; previously a research engineer at Meta AI, known for work in computer vision.

The three founders were schoolmates at Tsinghua, and the founding team carries an unusually research-heavy pedigree for a consumer-AI startup.

Competition and market position

Domestically, Moonshot is one of China’s “Six AI Tigers” and competes with DeepSeek, Alibaba’s Qwen, Zhipu (Z.ai / GLM), MiniMax and the big-tech assistants from ByteDance (Doubao) and Baidu (Ernie). On consumer reach it trails badly — ByteDance’s Doubao and Alibaba’s Qwen dwarf Kimi on monthly users — but on open-weight model quality it leads the pack, with Kimi K2.6 ahead of the next-best open model, Zhipu’s GLM-5.1.

Globally, Moonshot’s position is as the leading open-weight challenger rather than an absolute-frontier competitor. It does not try to out-scale OpenAI, Anthropic or Google; instead it gets within a few points of their frontier on independent benchmarks and gives the weights away, competing on cost and openness. That makes it central to the open-weight movement and to the US-China AI rivalry, and a natural counterpart to DeepSeek.

Controversies

• Distillation allegation (February 2026). Anthropic accused Moonshot, alongside DeepSeek and MiniMax, of using thousands of fraudulent accounts to harvest millions of conversations with Claude to train its own models. The claim is an allegation Moonshot has not accepted, and it has not been independently proven.
• Data residency. The Kimi consumer app stores data on servers in China, bringing it within scope of China’s National Intelligence Law — the same concern that applies to DeepSeek and Qwen, and the reason privacy-sensitive users are steered to self-hosting the open weights.
• US scrutiny of Chinese open models. Kimi’s use inside Western developer tools (including coding products) has drawn congressional attention in the US as part of a broader debate over Chinese open-weight models’ share of global downloads.
• Consumer decline. Kimi’s monthly active users fell sharply from a 2024 peak, and the open-weight pivot is in part a response to losing the consumer race at home.
• State-linked capital. The May 2026 round brought in state-backed China Mobile and Tsinghua Capital, sharpening questions — common to all Chinese AI labs — about independence and geopolitical exposure.

Recent developments (2026)

• ~$20 billion valuation — a roughly $2 billion round in May 2026, led by Meituan’s Long-Z Investments, the largest single financing in China’s LLM sector at the time.
• Hong Kong IPO preparation — Moonshot is reportedly dismantling its offshore (red-chip / VIE) structure to clear the path for a listing under China’s new IPO rules.
• Kimi K2.6 (April 2026) — became the leading open-weights model on independent testing.
• Kimi K2.7 Code and Kimi Work (June 2026) — a stronger agentic coding model and a local desktop agent extended Moonshot from chat into developer and knowledge-worker tooling.
• Distillation dispute — the February 2026 Anthropic allegation kept Moonshot in the centre of the data-sourcing debate around Chinese models.

Where Moonshot excels

• Open-weight quality. Kimi K2.6 is the strongest open-weights model on independent testing — frontier-adjacent capability anyone can self-host.
• Long context and agents. A founding focus on long context, now paired with the OK Computer agent mode and Agent Swarm, gives Kimi genuine depth for document and autonomous work.
• Cost. A low-priced API and free open weights undercut closed frontier labs, the same playbook that made DeepSeek a global story.
• Research infrastructure. Mooncake and the Muon optimiser show real engineering depth in doing more with constrained hardware.

Where Moonshot falls short

• Consumer reach. Kimi has lost the China consumer race to ByteDance’s Doubao and Alibaba’s Qwen, and has little Western consumer presence.
• Not the absolute frontier. Moonshot trails the top US labs on the hardest benchmarks; its pitch is “close and open”, not “best”.
• Jurisdiction and trust. China data residency and the distillation allegation make the hosted app a hard sell for regulated or sensitive use.
• Capital intensity. Competing at frontier scale under export controls is expensive, and reliance on big-tech and state-linked backers raises independence questions.

Developer resources

Moonshot’s developer stack centres on the Moonshot Open Platform (platform.moonshot.ai), a pay-as-you-go, OpenAI-compatible API for the Kimi K2 models, with context caching and agentic tool use. The Kimi K2 weights are published on Hugging Face under a modified MIT licence for self-hosting, and Kimi Code ships as a command-line coding agent. Consumer pricing for the Kimi app is on the membership page.

Frequently asked questions

Who owns Moonshot AI?

Moonshot AI is a private Beijing company founded in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin. Its major backers include Alibaba (reported to be its largest outside shareholder), Tencent, Meituan, HongShan, IDG Capital and, since May 2026, state-backed China Mobile. The company is restructuring to prepare for a possible Hong Kong listing.

How much is Moonshot AI worth?

Moonshot was valued at roughly $20 billion in a funding round in May 2026, when it raised about $2 billion led by Meituan’s Long-Z Investments. It is reportedly seeking a further round that would value it near $30 billion, though that figure is unconfirmed. As a private company, it does not publish audited financials.

What is Moonshot AI’s latest model?

The flagship chat model is Kimi K2.6, released in April 2026 and ranked the strongest open-weight model in the world by Artificial Analysis. In June 2026 Moonshot also released Kimi K2.7 Code, a coding-focused model, and Kimi Work, a local desktop agent.

Is Kimi the same as Moonshot AI?

Kimi is the product; Moonshot AI is the company that makes it. Kimi is both the consumer assistant app and the name of Moonshot’s model series (Kimi K2, K2.5, K2.6 and so on).

Are Moonshot’s models open source?

The Kimi K2 family is released as open weights under a modified MIT licence, so the models can be downloaded, fine-tuned and self-hosted. The licence requires attribution only for products with over 100 million monthly users or $20 million in monthly revenue. Self-hosting is the standard way to use Kimi outside China’s data jurisdiction.

How does Moonshot compare to DeepSeek?

Both are Chinese labs that release frontier-adjacent open-weight models cheaply, and both are central to the open-weight movement. DeepSeek offers a larger context window and a fully free app, while Moonshot’s Kimi K2.6 leads independent open-weight rankings and its OK Computer agent mode goes further on autonomous tasks. See the best AI models for a direct comparison.

Sources and official links

Models

Apps

• Kimi