Moonshot AI

Moonshot AI is a Chinese AI lab founded in 2023 by Yang Zhilin and headquartered in Beijing. It is one of China's "AI tigers," a cohort of well-funded domestic frontier-model startups that emerged in 2023. Its principal products are the Kimi chatbot family, the Kimi K2 open-weight Mixture-of-Experts model, and the Kimi K3 flagship released in July 2026.

Snapshot

Valuation

Revenue

Models and benchmarks

Overview

Moonshot AI rose to public attention with the Kimi chatbot, noted for early long-context capabilities (initially 200k tokens, later extended). In mid-2025 it released Kimi K2, a 1.04-trillion-parameter Mixture-of-Experts model with 32 billion active parameters, introducing the MuonClip optimizer (Muon plus QK-Clip) and a large-scale agentic data synthesis pipeline (Source: Kimi K2 Technical Report). Moonshot joins DeepSeek and Alibaba as Chinese labs publishing frontier-training research with optimization and efficiency contributions, and Kimi K2's training pipeline is a detailed public case of industrial-scale agent-trajectory synthesis (agentic AI).

People

Yang Zhilin (杨植麟) is the founder and CEO. He was previously a researcher at Google Brain and Carnegie Mellon and is a co-author on several widely cited NLP papers.

Models and technical approach

The Kimi chatbot launched in 2023 alongside the company's founding, and through 2024 Moonshot expanded its long-context capabilities. In July 2025 the company released the Kimi K2 technical report on arXiv, with weights under the CC BY-NC-ND 4.0 license (Source: Kimi K2 Technical Report). In April 2026 it released Kimi K2.6, which scored 58.6 on SWE-bench Pro at roughly 5–6× lower cost than Opus and demonstrated a 12-hour autonomous agent run (Source: artificialintelligencemade-simple.com).

On July 16, 2026, Moonshot launched Kimi K3, a ~2.8-trillion-parameter sparse MoE flagship with native vision and a 1-million-token context window, built on the Kimi Delta Attention and Attention Residuals architectures; the company describes it as "the world's first open 3T-class model," with self-reported benchmarks near Claude Fable 5 and GPT 5.6 Sol on agentic coding while trailing them overall, an API-first launch, and weights promised by July 27, 2026 (Kimi K3: Open Frontier Intelligence (Moonshot AI, July 2026)). Capability, pricing, reception, and market-reaction detail is on the model page. Demand after the launch nearly exceeded serving capacity, and Moonshot paused new subscriptions late on July 19, 2026 (Source: x.com; bloomberg.com).

Funding and IPO plans

Following the K3 release, Moonshot told investors it may pursue a Hong Kong IPO in as early as six months, reporting annual recurring revenue of $300 million in June 2026, up from $200 million in April; shares of domestic rivals Zhipu (Z.ai) and MiniMax fell 28.4% and 15.6% in the days after the launch (Source: bloomberg.com). On July 20, 2026 the company was reported to be seeking investor approval to begin the IPO process (Source: theinformation.com). A $30 billion funding round was separately reported forming in July 2026; dated valuation figures are in the Snapshot above. On July 21, 2026, Bloomberg reported the company in talks to raise pre-IPO funding at a $50 billion valuation, up from $31.5 billion at its last close, ahead of the Hong Kong listing (Source: bloomberg.com). The round closed at $3.5 billion on a $35 billion valuation, reported July 29, 2026, after an initial target of $1 billion to $2 billion; the company is approaching backers for a further round at a $50 billion pre-money valuation ahead of a possible Hong Kong listing this year (Source: finance.yahoo.com).

Several technical features of the K2 work are documented in the technical report (Source: Kimi K2 Technical Report):

• MuonClip optimizer. The K2 paper introduces QK-Clip, a per-head weight-rescaling stabilizer that eliminated loss spikes over 15.5T training tokens, a result for MoE training stability at trillion-parameter scale.
• Agentic training at scale. The pipeline used 3,000+ real MCP tools plus 20,000+ synthetic tools, stateful simulators, and LLM-as-judge filtering, one of the more detailed public accounts of industrial agentic data synthesis.
• Self-critique rubric rewards. A joint RL pipeline in which the critic is continuously refined by on-policy verifiable-reward rollouts.
• Long-context focus. The model halves the attention-head count (64 versus DeepSeek-V3's 128) to optimize 128k-context inference, described as an explicit design choice for agentic workloads.

Licensing

Kimi K2 weights are released under CC BY-NC-ND 4.0, which is research-only and bars derivatives and commercial use. This is more restrictive than peer Chinese open-weight releases such as DeepSeek's permissive licenses, and complicates classification as "open source" under the open-source AI frame; the K2 release is a case of partial openness in which weights are public but the license bars commercial derivatives.

Regulation

In April 2026 Beijing ordered Moonshot to reject U.S. capital without state approval, following Meta's acquisition of Manus (Source: bloomberg.com). The move situates Moonshot within US-China AI competition.

U.S. distillation and chip-access allegations

On July 22, 2026, White House OSTP Director Michael Kratsios said the administration has "information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," and separately said Moonshot acquired GB300-equipped servers and accessed the banned Nvidia chips in Thailand, likely for training (Source: x.com; bloomberg.com). In further remarks the same day, Kratsios accused Moonshot by name of building an internal platform for large-scale distillation against Anthropic's Fable model and of acquiring embargoed Nvidia GB300 servers to train Kimi K3 (Source: thehill.com). The Commerce Department's Bureau of Industry and Security confirmed the same day that it is formally investigating whether Chinese firms including Moonshot are accessing advanced U.S. AI chips, with Entity List designation a possible outcome (Source: theinformation.com). Treasury Secretary Scott Bessent said on July 21 that the U.S. "ha[s] the ability to sanction" overseas models that steal from American companies, citing "watermarks" of U.S. models found inside Chinese ones (Source: cnbc.com). Reporting on July 23, 2026 said the administration was reviving a push to ban Chinese AI models — which by then captured 46.4% of tokens routed through OpenRouter — following the K3 launch, over objections from administration figures including Sriram Krishnan and David Sacks, who argued a ban would hand OpenAI and Anthropic a domestic monopoly; the same coverage noted that downloadable open weights could make an outright U.S. ban difficult to enforce (Source: tomshardware.com; fastcompany.com). See Distillation, Chip Smuggling and Export-Control Evasion, Kimi K3.

A computing-power agreement between Moonshot and Alibaba Group Holding for the use of around 20,000 Nvidia chips became public on July 31, 2026. People with knowledge of the companies' operations described the arrangement as forming a substantial share of the capacity behind the Kimi models. The date the agreement was struck is not stated, and the source's body was behind a paywall with only the lede retrievable (Source: bloomberg.com).

Relationships

• related: DeepSeek, Alibaba / Qwen Team, China and the US Are Running Different AI Races
• related: Agentic AI, Open-Source AI / Open-Weight Models, Kimi K3