Moonshot AI reportedly closed a private financing round at a valuation of approximately $50B in October 2026, or roughly 50x its August annualized revenue, ahead of a planned Hong Kong IPO in the first quarter of 2027. Reuters reported in September 2026 that Moonshot had confidentially filed for a Hong Kong IPO, with the timetable still subject to regulatory approvals.

The company raised $3.5B at a reported $35B valuation in July 2026, following a $2B round at a $20B valuation in May.

Moonshot previously raised a $500M Series C at a $4.3B valuation in December 2025, led by IDG Capital with Alibaba and Tencent participating, following a $300M round in mid-2024 at roughly $3.3B. Alibaba led a $1B financing in February 2024 at a $2.5B valuation. Other backers include HongShan and 5Y Capital.

Yang Zhilin founded Moonshot in early 2023 after research work at Google and Meta and a teaching role at Tsinghua University. Launched in October 2023, Kimi differentiated through long context, processing up to 200,000 Chinese characters so users could analyze entire legal contracts or financial filings. Moonshot later expanded into coding and agentic knowledge work.

Kimi is Moonshot AI's family of AI models and applications for researching information, writing software and producing work from documents, images and video. Customers use Kimi through a web or mobile app, the Kimi Work desktop application, the Kimi Code coding tool or an API embedded in their own software.

A knowledge worker can point Kimi Work at a folder of quarterly reports and ask it to compare company performance. Kimi reads the documents, searches for additional information and runs code to analyze the data, then produces a summary, spreadsheet or presentation for the user to review. The desktop app can work with local files and schedule recurring tasks, such as preparing a daily briefing.

Kimi Chat handles conversations and document questions, while Deep Research gathers material for a longer investigation. Slides, Docs and Sheets turn that work into presentations, written documents and spreadsheets. Agent Swarm divides a larger assignment among specialized agents, and OK Computer carries out multi-step tasks using software tools rather than returning only a chat response.

Moonshot launched Kimi K3 in July 2026 as the world's largest open-weight model, with 2.8 trillion parameters, native understanding of images and video, and a context window of up to 1 million tokens. Its mixture-of-experts architecture activates 16 of 896 experts for a given input, allowing it to use a fraction of the full model on each step. Kimi Delta Attention and Attention Residuals improve how the model handles long inputs and passes information through its layers.

Kimi Work coordinates multiple specialized agents for research and office tasks, with integrated market data for Chinese, Hong Kong and US equities. Widgets can connect generated charts or interactive components to local data and external tools, while a dashboard keeps selected widgets in a persistent workspace.

Kimi Code runs in a developer's terminal and works with their repository, using command-line tools to inspect files, make changes and test software. K3 also uses screenshots as feedback when building interfaces, games and other visual software. Developers can call the same model through the Kimi API, while the released model weights let other organizations operate their own deployments subject to the Kimi K3 license.

Moonshot AI combines consumer and team subscriptions with usage-based API sales. Free access brings users into Kimi; paid plans sell greater access to its models and agent workflows, while enterprise offerings add account management and data privacy controls. Developers and businesses pay Moonshot to host models for their applications.

As of October 2026, international subscriptions cost $19 per month for Plus, $39 for Pro, $99 for Max and $199 for Ultra when billed monthly. Higher tiers sell larger usage allowances and more concurrent agent tasks across Kimi's applications, including Code, rather than a separate subscription for each output format.

API customers pay for input and output tokens, with different charges for cached and uncached input. Kimi K3 lists $3 per million uncached input tokens, $0.30 per million cached input tokens and $15 per million output tokens. Long documents, repeated agent steps and generated code all increase the amount of compute a customer consumes.

Moonshot's major costs are model training and GPU capacity for serving requests. Its Mooncake system separates parts of inference and caches reusable context to reduce repeated work; its sparse model architecture limits how much of K3 runs on each step. These efficiencies can improve serving economics, but a larger context window and longer agent tasks can still increase the compute needed per paying customer.

After spending roughly $100M on advertising in 2024, Moonshot cut ads and shifted toward distributing its models through businesses. The July 2025 release of Kimi K2's weights gave developers a model they could adapt to their own applications; Fireworks AI and Together AI now host standard or customized Kimi models, letting customers buy inference through providers with their own infrastructure and data-handling arrangements. Moonshot also sells through its applications and API, while enterprise engineers help organizations such as AsiaInfo Technologies and Kingsoft Cloud deploy its models.

Kimi K3 adds a separate licensing business. Companies operating a model-as-a-service business whose combined revenue with affiliates exceeds $20M over any consecutive 12 months must enter a separate agreement with Moonshot before commercial use. The model-as-a-service definition excludes end-user products with model capabilities embedded within specific features or agent workflows, while internal deployments are exempt from the separate-agreement requirement. Moonshot has signed revenue-sharing agreements with smaller platforms, while Reuters reported in August that it was seeking up to a 30% share of K3-related revenue from Microsoft, Amazon and Google; that percentage is a negotiating position, not a universal published rate or confirmation of signed hyperscaler agreements.

Moonshot competes directly with DeepSeek, Z.ai and MiniMax for developers and enterprises buying Chinese AI models, and with OpenAI and Anthropic in coding and knowledge work. Kimi K3 combines released model weights with hosted applications and an API, giving customers a choice between buying inference from Moonshot and running the model themselves.

Its central challenge is converting strong model performance into repeat paid usage while rivals compete on price, distribution and deployment costs. The customer can change the underlying model more easily than a deeply integrated business application, so coding tools and persistent workspaces matter to retention.

DeepSeek competes through low-cost reasoning and coding models, while Z.ai and MiniMax also sell model APIs and agent capabilities. Where Chinese labs have emphasized smaller models that fit constrained GPU capacity, Moonshot bet on scale: K3's 2.8 trillion parameters compare with 763 billion for DeepSeek V4.1 Flash and roughly 743 billion for GLM-5.3. Larger weights can support more capable specialized agents, but require more memory and infrastructure to serve, making international inference providers important to distribution.

OpenAI and Anthropic compete for the same international developer and knowledge-worker budgets through their assistants, APIs and coding agents. Companies building specialized coding and legal agents can adapt Kimi's released weights and control where inference runs rather than depending entirely on a closed-model supplier. K3's July benchmark advantage has already been overtaken by newer models from Xiaomi, Alibaba, DeepSeek and Z.ai, increasing the importance of Moonshot's applications and customer workflows beyond any one model release.

Alibaba, Baidu, ByteDance and Tencent can distribute their own models through cloud services, enterprise software and consumer applications. Free assistants from ByteDance, Alibaba and Tencent squeeze consumer pricing, while longer context windows from Alibaba and Baidu have eroded Kimi's original differentiation. Alibaba's Qwen competes with Kimi even though Alibaba also backs Moonshot; a cloud customer can buy a bundled model without establishing a separate relationship with an independent lab.

Cloud partners could broaden Moonshot's reach, but they also control customer access and offer competing models. Moonshot's proposed K3 revenue-sharing arrangements seek payment when those platforms sell its model rather than leaving the infrastructure provider to capture the entire hosted-service bill.

Moonshot is expanding from individual users and developers into enterprise deployments and paid distribution through external cloud platforms. Both opportunities depend on turning adoption of its models into spending by buyers who need deployment support, organizational controls or a managed service.

Moonshot has assembled engineers to help customers including AsiaInfo Technologies and Kingsoft Cloud deploy its models. Kimi Work and enterprise account controls give it a way to sell into knowledge-work budgets beyond consumer chat, but larger contracts require reliable integrations, data handling and support alongside model quality.

Microsoft Azure, Amazon Web Services and Google Cloud could put K3 in front of enterprises that already procure AI through a cloud provider. The separate commercial-license requirement gives Moonshot a basis for negotiating payment from large hosts; signed distribution and revenue-sharing agreements would determine how much of that demand it can monetize.

Beyond its Chinese user base, Moonshot is targeting knowledge workers in markets including South Korea and Brazil, using local-language influencer marketing to reach new customers. Frontier model capabilities and a broad productivity subscription compete with domestic models such as Naver's HyperCLOVA X and Maritaca's Sabiá, whose local infrastructure and security positioning can matter more to regulated buyers. Winning those budgets requires localization and trusted deployment options alongside benchmark performance.

Compute economics: Moonshot recommends deploying K3 on configurations with at least 64 accelerators, and its agent workflows can run long sequences of inference calls. Lower token prices or heavier use by subscribers could reduce the revenue retained after serving costs even as annualized revenue rises.

Model demand and distribution: DeepSeek’s January 2025 R1 release helped reverse Kimi’s early consumer lead, and K3’s benchmark advantage proved temporary, exposing Moonshot to customers changing models before it recovers training costs. Its commercial license requires large model hosts to negotiate, while the reported 30% cloud revenue share remains a negotiating position that could make competing models more attractive.

Model provenance and regulatory scrutiny: Anthropic accused Moonshot of unauthorized distillation in September 2026, and The Information reported in October that OpenAI’s Astra outputs helped develop Kimi K3.1 Pro. These allegations, which Moonshot had not publicly addressed, could lead to tighter model access and disputes that complicate enterprise adoption and its planned Hong Kong IPO.