Moonshot AI

Moonshot AI is a Beijing-based artificial intelligence company that develops the Kimi chatbot, large language models, and agent software. The company says it was founded in early 2023. Contemporary reporting identifies Yang Zhilin, Zhou Xinyu, and Wu Yuxin as its co-founders, with Yang serving as chief executive; Chinese business press also names Zhang Yutao as a co-founder and chief technology officer.^{[1][2][43][78]} Its principal operating entity, Beijing Moonshot AI Technology Co., Ltd. (北京月之暗面科技有限公司), converted to a joint stock limited company on July 29, 2026 as part of its listing preparations and was renamed 北京月之暗面科技股份有限公司.^[1][63]

Moonshot first became widely known for Kimi's long-document handling. It later shifted much of its technical identity toward downloadable model weights, sparse mixture-of-experts architectures, reasoning, coding, tool use, and multimodal agents. The public model line includes Kimi K1.5, Kimi K2, Kimi K2 Thinking, Kimi K2.5, Kimi K2.6, Kimi K2.7 Code, and Kimi K3. As of September 23, 2026, K3 remained the flagship; on September 11, Moonshot rolled out a Kimi K2.8 Preview model in Kimi Code.^[4][28][57]

Moonshot describes its long-term objective in terms of artificial general intelligence.^[1] That is an organizational aspiration, not evidence that its current systems are generally intelligent. The company's models remain probabilistic generative AI systems whose capabilities and limitations vary by checkpoint, interface, tool configuration, and deployment.

History

Founding and technical background

Moonshot was formed in Beijing in early 2023; Chinese registry reports give April 2023 as the date the operating company was established.^[1][63] Its Chinese name, Yue Zhi An Mian, means "the dark side of the moon", and profiles of Yang connect the name to the Pink Floyd album, his favorite.^[1][2]

Yang studied at Tsinghua University and earned a PhD from Carnegie Mellon University in 2019. His academic page records research appointments at Google Brain and Meta AI. He co-authored Transformer-XL and XLNet, two influential sequence-modeling papers built on the transformer architecture.^[6][7] In July 2026, Moonshot's English company page said its core technical team included the inventors of Transformer-XL, rotary position embedding (RoPE), Group Normalization, ShuffleNet, MuonClip, and Mooncake.^[1] This is a team-level description and should not be read as assigning every invention to a founder.

Reporting identifies Zhou and Wu as co-founders and fellow Tsinghua University alumni.^[43] A 2024 report by the Information Technology and Innovation Foundation says Zhou previously worked at Hulu, Tencent, and Megvii, and that Wu worked at Google Brain on foundation models and at Meta AI Research on computer vision.^[8] The South China Morning Post reported that Yang assembled a team of about 40 AI specialists in the company's first three months.^[43] Because public biographies for the two co-founders are limited, the article does not infer their current duties from older employment histories. A fourth co-founder, Zhang Yutao, is described in Chinese business press as chief technology officer and is one of two individual respondents in the Recurrent AI arbitration described below.^[76]

Kimi launch and long context

Moonshot announced Kimi Chat in October 2023 with support for prompts of about 200,000 Chinese characters. The service opened more broadly in November after an invitation-based testing period.^[9] In March 2024, Moonshot announced an invited beta that raised the input limit to about two million Chinese characters.^[10] These figures described particular product releases. Chinese characters and model tokens are different units, and later interfaces have used model-specific token limits.

The long-context product depended on serving infrastructure as well as a model checkpoint. Moonshot's Mooncake system separates prompt processing from token generation and treats the key-value cache as a distributed resource. The work received a Best Paper award at the 2025 USENIX Conference on File and Storage Technologies. Its authors reported operation across thousands of nodes and more than 100 billion processed tokens per day, along with higher request capacity than their comparison systems on Kimi traces.^[11] Those capacity figures are author-reported production results, not an independent comparison of answer quality.

From hosted chat to open-weight models

Moonshot initially distributed its strongest capabilities mainly through the Kimi service and API. In 2025 it began publishing more model weights, code, and technical reports. Kimi K1.5 focused on long-context reinforcement learning for multimodal reasoning. Kimi K2 introduced a one-trillion-parameter sparse architecture and was released with base and instruction checkpoints. Subsequent K2-series releases added longer contexts, persistent tool use, visual input, and multi-agent orchestration.^{[12][13][20][21][24]}

This shift made Moonshot more visible outside China. It also created an important terminology distinction. A downloadable checkpoint is open-weight, but it is not necessarily open source in the sense of publishing the complete training data, data-processing pipeline, and unrestricted licensing. Moonshot's K2 and K3 releases use different licenses, and commercial deployers must review the terms attached to the exact model.^[14][15]

The K3 release as a company event

Main article: Kimi K3.

K3 was a Moonshot launch that moved capital markets, regulators, and the company's own capacity planning within a fortnight. The sequence is short and well dated. Moonshot announced K3 on July 16, 2026 and made it immediately available through its Kimi app, Kimi Work, Kimi Code, and the API, while promising full weights by July 27.^[4][28] It met that date, publishing the weights, custom modeling code, technical report, and supporting infrastructure on July 27.^[27][45][66]

Demand outran provisioning almost at once. Moonshot paused new consumer subscriptions on the evening of July 19 and gave its available computing power to existing paid users; Reuters reported the next day that the company also planned to split future memberships into two plans, one of them only for coding.^[39][59][68] Chinese coverage of Bloomberg's reporting said daily sales rose at least sixfold after the launch, attributed to a person familiar with the matter.^[59][83]

The commercial terms changed as well. Where K2.5, K2.6, and K2.7 Code use a Modified MIT license, K3 shipped under a bespoke Kimi K3 License whose model-as-a-service and attribution conditions are keyed to a deployer's revenue and user counts.^[15][25][55] That is a company decision about how far the open-weight strategy extends, not a technical property of the checkpoint.

The reception repositioned Moonshot among Chinese laboratories. Caixin reported that K3 ranked third on the Artificial Analysis index behind Claude Fable 5 and GPT-5.6 Sol and first for coding on Code Arena, and that at 2.8 trillion parameters it displaced DeepSeek-V4-Pro as the largest publicly released open-weight model.^[66] Writing on the Interconnects newsletter on July 20, Nathan Lambert placed K3 second on the Vals AI index and third on the Artificial Analysis Intelligence Index and described Moonshot as going "toe to toe with Anthropic and OpenAI with far, far fewer resources".^[74] Both are dated snapshots of moving leaderboards rather than settled rankings.

Products and services

Kimi

Kimi is Moonshot's hosted assistant for chat, web search, file analysis, multimodal input, writing, coding, and research. Current documentation describes web and mobile clients, a developer API, Kimi Code, Kimi Work, and agent modes that can create websites, documents, spreadsheets, and presentations.^[16] Product availability, context limits, pricing, and membership rules can change independently of a model release.

Moonshot introduced an agent mode called OK Computer in September 2025. The current Kimi Agent incorporates web development, document generation, data analysis, Deep Research, Kimi Slides, and Agent Swarm workflows.^[17] Kimi Work, launched in beta on June 3, 2026, brings agent execution to local desktop workflows and can use local files and browser automation when the user grants access.^[18] These products are tool-using applications around Kimi models; they are not separate neural-network architectures.

The service can be useful for long documents and multi-step tasks, but long context does not guarantee that every passage will be retrieved or interpreted correctly. Agent access also adds operational risks such as prompt injection, excessive permissions, destructive tool actions, and unsupported conclusions. Important deployments require source checking, least-privilege access, and review of generated actions.^[44]

By mid-2026 the surface had widened again. The help center distinguishes Kimi, Kimi Work, Kimi Code, and Kimi Claw, a zero-deployment cloud service for continuously running agents that entered public beta in February 2026; in Kimi Code and Kimi Work, a component called WebBridge lets the agent operate a browser. The help center also documents memory, Projects, and features for websites, slides, documents, spreadsheets, and Deep Research, and offers K2.6, K3, and a K3 Swarm mode as model choices. Agent Swarm is documented as coordinating up to 300 sub-agents in parallel.^[16][17][18] A July 2026 version of Moonshot's English company page added Kimi Business and claimed Kimi had "tens of millions of professional users monthly", a company figure with no published methodology.^[1]

Kimi Code and the command-line agents

Moonshot maintains its coding agent as open-source software rather than as a closed client. Kimi CLI, an Apache-licensed terminal agent with a shell mode, a Visual Studio Code extension, and Agent Client Protocol support for other editors, appeared in October 2025.^[45][53] In May 2026 the company started a successor, Kimi Code CLI, distributed as a single binary under the MIT license with video input, conversational Model Context Protocol configuration, and a plugin marketplace.^[53] By September 2026 the kimi-cli repository had been archived; its README says the Python Kimi CLI has been replaced by Kimi Code CLI, will receive no further releases or security updates, and that existing installations will stop working.^[53] On September 17, 2026, Moonshot released version 2.0 of Kimi Code CLI together with a Kimi Code desktop app for macOS and Windows that puts the same agent core behind a graphical interface.^[28] On September 23, 2026, the GitHub API counted about 11,400 stars for kimi-cli and 7,600 for kimi-code, against about 11,100 for the Kimi-K2 repository.^[45] Star counts measure attention, not deployment.

Developer platform

Moonshot offers an API for chat, reasoning, vision, and agent-oriented models. Its documentation uses request patterns compatible with common chat-completions clients, but compatibility does not imply identical behavior to OpenAI or another provider. Model identifiers, supported parameters, caching rules, and context limits remain provider-specific.^[19]

The billing help pages describe pricing per token, with input and output tokens priced separately, prices that differ by model, and an additional charge for each web search; the platform also documents a context-caching feature.^[3][57] Because prices are operational data rather than stable model properties, this article does not reproduce a static price table.

The platform's public model list is short. As of September 23, 2026 it offered kimi-k3 at a one-million-token context, kimi-k2.7-code-highspeed and kimi-k2.6 at 256,000 tokens, with the K2.7 entry positioned for coding work that needs higher output speed.^[57] Moonshot also publishes a vendor verifier, a test harness that checks whether third-party hosts of its models reproduce the reference behavior.^[45] Its K3 results table lists submissions from inference providers including Fireworks, Baseten, Together, Nebius, and Modal alongside Moonshot's own endpoint.^[95]

Model family

The table lists major public releases. Dates refer to public introductions or official model documentation, not necessarily the later date of a technical paper.

Kimi k1.5

The Kimi k1.5 technical report describes a multimodal model trained with long-context reinforcement learning. Its framework uses partial rollouts and methods for shortening long reasoning traces, without relying on Monte Carlo tree search, a learned value function, or process reward models. The authors reported strong mathematics, coding, and visual-reasoning results.^[12] The scores are developer-run experiments. The report does not state a parameter count, so the frequently repeated estimate of about 500 billion parameters has no support in it.

Kimi K2 and K2 Thinking

Kimi K2 uses a sparse mixture-of-experts design with one trillion total parameters and 32 billion activated parameters per token. The technical report describes 61 layers, 384 experts with eight selected per token, a 128,000-token context in the initial release, and pretraining on 15.5 trillion tokens. It also introduces MuonClip, which the authors used to train the model without a reported loss spike.^[13] Moonshot released base and instruction checkpoints with configuration and inference code.^[14]

The September 2025 K2 Instruct 0905 update increased the documented context from 128,000 to 256,000 tokens and targeted coding and agent tasks.^[20] K2 Thinking followed on November 6. Moonshot presented it as a model that could continue reasoning while making long sequences of tool calls.^[21] A December 2025 evaluation by the United States Center for AI Standards and Innovation (CAISI) called K2 Thinking the most capable model from a PRC-based developer at its release, but found it only a modest improvement on DeepSeek V3.1 and below leading US models, including older ones such as GPT-5 and Claude Opus 4, on agentic cyber and software-engineering tasks. The evaluator also found the model highly censored in Chinese and relatively uncensored in English, Spanish, and Arabic.^[22]

Public discussion sometimes repeated a US$4.6 million training-cost estimate for K2 Thinking as if Moonshot had disclosed it. Yang said that figure was not official and did not represent the company's full cost.^[23]

Kimi K2.5 and K2.6

Kimi K2.5 added native image input to the K2 line and emphasized joint text-vision training. Its report says continued pretraining used about 15 trillion mixed visual and text tokens. It also describes Agent Swarm, an orchestration method that delegates tasks to parallel subagents. The report says Agent Swarm cut latency by up to 4.5 times against single-agent baselines on selected tasks, and Moonshot later described the K2.5 version as supporting up to 100 sub-agents and 1,500 steps.^[24][26] Those are results for the reported orchestration setup, not a general guarantee of reliability or speed.

Moonshot published K2.5 weights and code. Its modified MIT license adds a display requirement for commercial products or services above either 100 million monthly active users or US$20 million in monthly revenue.^[25]

Kimi K2.6 was introduced on April 20, 2026. Moonshot positioned it for long-horizon coding and tool use and documented an Agent Swarm configuration of up to 300 subagents and 4,000 coordinated steps.^[26] These values are system limits reported by the developer. They do not mean that every task needs that many agents or that the system will complete an arbitrary workflow correctly.

Kimi K2.7 Code

Kimi K2.7 Code was released on June 12, 2026 as a coding-specialized member of the K2 line rather than a new backbone. Its model card describes the same K2 shape of one trillion total and 32 billion activated parameters with a 256,000-token context, released under the Modified MIT license.^[55][56] Moonshot reports gains over K2.6 on its own suites, including Kimi Code Bench v2 rising from 50.9 to 62.0, MLS Bench Lite from 26.7 to 35.1, and MCP Mark Verified from 72.8 to 81.1, together with about 30 percent fewer thinking tokens on comparable tasks.^[55] Kimi Code Bench v2 is Moonshot's in-house benchmark, and all of these scores come from Moonshot's own runs, so they describe developer measurement rather than an independent comparison. The API exposes the checkpoint as kimi-k2.7-code-highspeed, and it remained the recommended option for speed-sensitive coding work after K3 shipped.^[57]

Kimi K3

Moonshot made K3 available through Kimi, Kimi Code, Kimi Work, and its API on July 16, 2026, and published the full weights through GitHub and Hugging Face by July 27.^[4][27] The model card describes 2.8 trillion total parameters, 104 billion activated parameters per token, 93 layers, 896 routed experts with 16 selected per token, a native visual encoder (the 401-million-parameter MoonViT-V2), and a context window of 1,048,576 tokens. Its attention stack combines 69 Kimi Delta Attention layers with 24 gated multi-head latent attention layers.^[27]

Moonshot's launch material says K3 applies quantization-aware training from the supervised fine-tuning stage onward, with MXFP4 weights and MXFP8 activations, and reports an approximately 2.5-fold improvement in overall scaling efficiency over K2. It also states that the model still trailed the then-current Claude Fable 5 and GPT-5.6 Sol in the company's overall comparison.^[4] Both the efficiency figure and benchmark table are developer results. They combine different reasoning budgets, tools, and evaluation harnesses and should not be converted into a categorical ranking.

The K3 release attracted more hosted demand than Moonshot had provisioned. On the evening of July 19, the company temporarily paused new consumer subscriptions while prioritizing existing paid users.^[39][59] That capacity decision does not establish anything about model accuracy.

K3 uses a bespoke Kimi K3 License rather than the K2 modified MIT terms. The license permits use, modification, and redistribution subject to conditions. It requires certain model-as-a-service operators whose group revenue exceeds US$20 million over a consecutive 12-month period to obtain a separate agreement. It also imposes prominent attribution requirements on products above specified user or revenue thresholds, with stated exceptions.^[15] "Open-weight" is therefore more precise than "unrestricted open source."

Research and technical contributions

Mooncake

Mooncake is Moonshot's distributed serving architecture for Kimi. It disaggregates prefill and decoding and uses CPU memory, storage, and networking resources to build a distributed key-value cache. The 2025 USENIX paper reported 59 to 498 percent more effective request capacity than baseline methods on its evaluated traces while meeting the authors' service-level objectives.^[11] The paper's deployment figures, 115 and 107 percent more requests handled on Nvidia A800 and H800 clusters than previous systems, are likewise first-party systems results.

The code is not in Moonshot's own GitHub organization. Mooncake is published under kvcache-ai, an open-source organization that describes itself as a collaboration between MADSys and industry collaborators, with the repository created on June 25, 2024 under the Apache 2.0 license; the Mooncake paper's authors are from Moonshot AI and Tsinghua University.^[11][48]

Muon and Moonlight

Moonshot researchers adapted the Muon optimizer for large language-model training by adding weight decay and controlling per-parameter update scale. Scaling-law experiments in their paper reported about twice the computational efficiency of AdamW under the tested compute-optimal setup.^[29] That is an experimental comparison with a defined training recipe, not a universal result against stochastic gradient descent or every AdamW configuration.

The same work introduced Moonlight, a sparse model with 3 billion activated and 16 billion total parameters trained on 5.7 trillion tokens. The authors released pretrained, instruction-tuned, and intermediate checkpoints along with a distributed Muon implementation.^[29] Moonlight served as a research vehicle for the optimizer rather than a replacement for the consumer Kimi model line.

K2 later used MuonClip, a related training method that adds QK-Clip to control attention logits at trillion-parameter scale.^[13] The K2 report's statement that the run had zero loss spikes describes that training run; it does not prove that MuonClip prevents instability in every configuration.

Kimi Linear

Kimi Linear is a hybrid linear attention architecture built around Kimi Delta Attention and multi-head latent attention. The 48-billion-parameter research model activated 3 billion parameters per token. Moonshot's experiments reported up to 75 percent less key-value-cache use and up to six times decoding throughput at a one-million-token context compared with the full-attention setup used in the paper.^[30] The team released kernels, inference implementations, and checkpoints. These are author-reported efficiency results and depend on model shape, hardware, sequence length, and software.

The architecture proved to be a rehearsal rather than a side project. Kimi Linear interleaves three KDA layers with one full MLA layer; K3 carries the same hybrid to 2.8 trillion total parameters with 69 KDA layers and 24 gated MLA layers, close to that three-to-one ratio.^[27][30]

Vision, coding, and agent research

Kimi-VL is a sparse vision-language model introduced in April 2025. Its report describes a 128,000-token context, a native-resolution MoonViT image encoder, and a language decoder with 2.8 billion activated parameters.^[31] Moonshot released its code and checkpoints, including a later thinking variant.

Kimi-Dev-72B is an open coding model intended for software-engineering tasks. Moonshot's repository reports a 60.4 percent result on SWE-bench Verified for its original agent setup.^[32] That score belongs to the documented harness and model revision. It should not be compared directly with scores from SWE-bench runs using different repositories, tools, token budgets, or patch-validation rules.

Moonshot also publishes evaluation work of its own. PerceptionBench, released on July 24, 2026, is a 3,000-question benchmark built by diagnosing the earliest failure points of frontier multimodal models across 42 existing benchmarks and reducing them to ten atomic perceptual skills. Its reported leaderboard puts no model above 60 percent, with GPT-5.6 Sol at 59.7 and K3 second at 58.5.^[50] The benchmark is Moonshot's own, so its ranking of Moonshot's model is a developer result.

Moonshot's technical releases span model training, serving, vision, coding, and AI agents. They also show that "model" and "system" are different evaluation units. Agent Swarm results include an orchestration layer, tools, parallel inference, and stopping rules in addition to the underlying neural network.

Open-source releases and infrastructure

By 2026 Moonshot was publishing training and serving infrastructure, not only weights. The pattern matters for how the company competes: the parts that are hardest to reproduce, such as kernels, communication libraries, and serving architecture, are released permissively, while the flagship weights carry the more restrictive Kimi K3 License. The table lists the main public repositories, with dates taken from the GitHub API and licenses as GitHub identifies them, except where a repository's own license file or README is cited.^[45]

Three details that have been attached to the July 2026 releases, including in some press coverage, do not survive checking. First, FlashKDA was not new: the repository has been public since April 20, 2026, three months before the K3 launch.^[45][47] Second, AgentENV, the Firecracker-based sandbox platform used for K3's agentic reinforcement learning, belongs to the kvcache-ai organization, not to MoonshotAI; a full listing of Moonshot's 43 public repositories contains no agent-environment project, and the only repositories the organization created in July 2026 are Kimi-K3, MoonEP, PerceptionBench, minitriton, and nano-kpu.^[45][49] Third, the roughly 2.5-fold improvement in scaling efficiency over K2 that Moonshot reports is attributed in its own launch post to structural changes in the model, namely Kimi Delta Attention, Attention Residuals, and the Stable LatentMoE routing framework, together with a refined data and training recipe. It is not a MoonEP result.^[4][46]

MoonEP is the genuinely new infrastructure release of the group. It is an expert-parallel communication library that guarantees every rank receives exactly the same number of tokens no matter how skewed the router output becomes, by planning a bounded set of duplicated experts online at each step. The repository's first commit is dated July 24, 2026 and the announcement July 27; its comparisons against DeepEP v2, run on H20 GPUs with eight-way expert parallelism, are published as plots rather than tables.^[46]

Two of the July repositories are of a different kind. minitriton is a tile compiler that lowers a Python-embedded kernel DSL through MLIR to PTX, with an eager tensor library on top; nano-kpu is the register-transfer-level design of a small hybrid-architecture inference chip together with a simulation and area-and-timing flow. The minitriton README says the whole stack was designed, implemented, measured, and written by K3 "with engineering direction and review by the maintainer", and the nano-kpu README says the chip was "designed and implemented fully by Kimi-K3"; both carry disclaimers that they are demonstrations of the model rather than Moonshot products.^[51][52] Moonshot's launch post says K3 completed the chip design in a single 48-hour run, closing timing at 100 MHz and sustaining more than 8,700 decode tokens per second in simulation.^[4] Those are simulation results on the open Nangate 45-nanometer cell library, which Investing.com noted is several generations behind the 3- and 2-nanometer nodes used for leading AI accelerators; the published artifacts are RTL and a simulation and timing flow.^[52][69]

The published weights do get taken up. On September 23, 2026, the Hugging Face API reported about 1.86 million downloads for the K3 repository, against roughly 441,000 for K2.6, 313,000 for K2.5, and 105,000 for K2.7 Code.^[75] The Hub counts a rolling window and does not distinguish an evaluation pull from a production deployment, so the figures show interest rather than installed base.

Funding and ownership

Moonshot is privately held, so most financing terms outside public-company filings come from people familiar with private transactions or from advisers. Amounts and valuations in the table are attributed reports, not audited company accounts.

Alibaba's annual filing is the strongest public evidence for a specific ownership figure. It says Alibaba Cloud's parent invested about US$800 million for an approximately 36 percent preferred equity interest during fiscal 2024.^[34] The company repeated the same wording in its annual report for the fiscal year ended March 31, 2026, filed on May 20, 2026, which describes the investment in "Moonshot AI Ltd" as accounted for under the measurement alternative and does not restate a current percentage.^[62] Several financing rounds have followed, so the 36 percent figure should not be read as Alibaba's stake today.

Chinese state-linked capital entered the register during 2026. In May 2026, STAR Market Daily reported, as relayed by IT Home, that Guozhitou, the Beijing Artificial Intelligence Fund, and the state-owned carrier China Mobile had joined Moonshot's shareholder register.^[65] In July 2026, Cailian Press reported that a National Social Security Fund technology-innovation vehicle and other institutions appeared in the company's business registration, and that registered capital had risen from RMB 1 million to about RMB 1.52 million; a person familiar with the matter said the entry reflected a 2025 investment whose registration had been completed late rather than a new round.^[64]

The July figures superseded an earlier, more tentative set. In June and July 2026, reports said Moonshot was raising again and considering a Hong Kong initial public offering. Reuters wrote on July 20 that the company was seeking up to US$2 billion in new capital and preparing for a potential listing. Bloomberg reported on July 21 that a summer round at a US$31.5 billion valuation was expected to close in the coming days and that a later pre-IPO round might seek a valuation as high as US$50 billion.^[39] Those were pending transactions when reported. Bloomberg then reported on July 29, according to Chinese relays of its story, that the round had closed at about US$3.5 billion, well above the original US$1 billion to US$2 billion target, at a US$35 billion post-money valuation, which is consistent with the US$31.5 billion pre-money figure.^[59][83] Chinese reports added that the oversubscription had led Moonshot to close the round ahead of schedule and that the follow-on G round, originally planned to open in August 2026, had been brought forward at a US$50 billion pre-money valuation.^[59][80] These remain accounts of private transactions attributed to unnamed people familiar with them, not published terms.

Corporate restructuring and listing plans

Moonshot spent 2026 rebuilding itself into something that can list. Bloomberg reported on May 19, 2026 that Moonshot had emailed shareholders that week to say it would begin dismantling its red-chip structure, the arrangement in which an offshore holding company controls the mainland business, to win approval for a Hong Kong listing; people familiar with the plan said a joint-venture structure would let US-dollar funds stay invested.^[60] On July 19, Cailian Press and Bloomberg reported that Moonshot had sent investors a proposal to list in Hong Kong, with a listing expected within about six months at the earliest; Reuters reported the next day that the company had engaged Goldman Sachs and China International Capital Corporation as advisers on the plan.^[39][60][61] KrASIA, adapting the Chinese outlet IPO Zaozhidao, wrote on July 30 that the two banks would act as joint underwriters, a reported arrangement rather than a company announcement.^[82]

The reported timetable then drew a direct company response. On August 3, 2026, Chinese outlets including Cailian Press and Sina Technology relayed a report that Moonshot planned to submit its Hong Kong listing application as early as that month and to raise about US$3 billion; the relays did not name an original source, and one quoted a person close to the company questioning whether Moonshot would seek less in a listing than the more than US$3.5 billion it had just raised privately.^[80] The same day, Moonshot told Jiemian News that the report was untrue, and a person familiar with the matter gave National Business Daily the same answer; neither response specified whether the timing, the amount, or both were disputed.^[81] The denial applied to that day's report. The shareholder proposal, the adviser discussions, and the July 29 conversion to a joint stock company are separately documented parts of the listing preparation.

A month later the filing itself was reported. On September 3, 2026, Reuters, citing three people with knowledge of the plans, reported that Moonshot had confidentially filed for a Hong Kong initial public offering, a filing first reported by LatePost; Moonshot told Chinese media it would not comment on market rumors and had nothing to disclose.^[5][84] Reuters added that, to gain regulatory approval, Moonshot had unwound its red-chip structure and moved to an onshore Chinese domicile before filing, and that it was in talks with Microsoft, Amazon, and Google on revenue-sharing agreements that would let the US cloud companies host its models.^[5] On September 12, after an unsourced chat screenshot about the company spread through Chinese investor groups, Moonshot's legal department said on Weibo that online claims about its founders and employees were fabricated and that it had reported them to the police.^[93]

The registry caught up on July 29, 2026. Chinese outlets citing Tianyancha business-registration data reported that the operating entity had converted from a limited liability company to a joint stock limited company, a change carried through into its registered Chinese name, with registered capital of about RMB 1.52 million. Registered capital in China is a nominal subscription figure and is unrelated to the money actually raised. Yang Zhilin moved from director to chairman and general manager, Zhou Xinyu from general manager to director, Zhang Yutong was added as a director, and Song Sijia was recorded as head of finance.^[63] Conversion to a joint stock company is a standard precondition for a Chinese issuer preparing to list; it is not itself evidence that a listing will happen or when.

Revenue figures are now reported often enough to record, with the caveat that annualized run-rate is an extrapolation from a short window and Moonshot has published no audited accounts. Reported annual recurring revenue passed about US$100 million in March 2026 and US$200 million by April or May (sources differ), and exceeded US$300 million by mid-June; Shanghai Securities News reported that API sales accounted for more than 70 percent of revenue.^[38][60][87] On September 11, Bloomberg reported that Moonshot was targeting annualized revenue of US$2 billion by the end of 2026, which TechCrunch described as double its reported August run rate; Chinese coverage of the same report put August ARR above US$1 billion.^[86][93] Those are attributed, unaudited figures. They describe a business whose growth is coming from developers rather than from the consumer app.

Competition and market position

Chinese media describe Moonshot as one of the "AI Six Tigers", a label for venture-backed foundation-model startups whose roster varies by source and date.^[93] Coverage commonly compares it with DeepSeek, MiniMax, and Zhipu AI, while major internet companies operate Qwen, Doubao, Ernie, and other model families.^[38][68] The label is a market narrative, not a technical classification.

Two of that group reached the public market first. Zhipu AI listed in Hong Kong on January 8, 2026 and MiniMax on January 9, each seeking in the region of half a billion US dollars, and both offerings were oversubscribed. Both companies were also loss-making at the time of listing.^[71] Other members of the cohort, including StepFun, were also preparing Hong Kong listings in 2026.

Moonshot also competes for developers and users with ChatGPT, Claude, Gemini, and Llama. Comparisons with Anthropic, DeepMind, and other laboratories vary by model version and task. Self-reported launch tables often mix hosted models, open-weight checkpoints, tools, and different inference budgets.

Third-party services such as Artificial Analysis and OpenRouter can help compare availability, latency, price, or selected evaluations, but their results remain snapshots of particular endpoints and settings. Benchmarks such as GPQA Diamond, Humanity's Last Exam, BrowseComp, and GDPval measure different tasks. A model's result on one does not establish general superiority.

Consumer reach versus developer reach

Moonshot's technical standing and its Chinese consumer position have moved in opposite directions. Reuters reported in July 2025 that Kimi had fallen from third place by monthly active users in August 2024 to seventh in June 2025 on the tracker aicpb.com.^[33] The National Business Daily's quarterly AI application ranking, drawing on third-party app measurement, put Kimi's monthly active users at 8.338 million in the first quarter of 2026, a fourth consecutive quarterly decline, with average monthly downloads of 2.506 million, down 11.2 percent from the previous quarter. The analysis read the fall as a deliberate reallocation toward model capability, developer ecosystem, agent products, and enterprise scenarios rather than a fight for consumer app rankings.^[72] QuestMobile's half-year report, published July 14, 2026, shows the same shape from the other side: it put the June 2026 active users of the three leading AI-native apps, Doubao, Qwen, and DeepSeek, at 382 million, 167 million, and 129 million, and its published summary mentions Kimi only in an engagement comparison, with 26.1 percent of Kimi users in its more-than-ten-minutes usage band.^[73] The API-led revenue mix reported by Shanghai Securities News is consistent with that picture.^[87]

Market reaction to K3

The K3 launch coincided with a sharp fall in semiconductor and AI-related equities, and outlets divided over how much of it to attribute to Moonshot. Shares in the electronic design automation vendors Cadence Design Systems and Synopsys fell about 9 percent on July 17 after Moonshot's claim that K3 had produced a chip design using only open-source tools, with the Nasdaq composite down about 1 percent; Bernstein's Robin Zhu said the episode was another case where the ability of China's top AI laboratories to keep pace with the US frontier had surprised global investors.^[69] A same-day analysis argued the selloff was overdetermined, listing disappointing Netflix and TSMC results, geopolitical risk, and rate fears alongside K3, and framed the underlying anxiety as a question about whether roughly US$700 billion of annual hyperscaler infrastructure spending can be repaid if capable models become cheap.^[70] Neither account establishes causation, and both name other drivers active in the same week.

Evaluation, safety, and scrutiny

Benchmark interpretation

Moonshot publishes extensive model tables against systems from OpenAI, Anthropic, Google, and other developers. These tables are useful when the prompt templates, reasoning effort, tools, sampling, and harness are disclosed. They are not stable league tables. For example, results for GPT-4, the OpenAI o-series, GPT-5.5, Claude Opus 4.8, and Claude Fable 5 may come from different access dates or operating modes.

Company-reported results are therefore best read as developer measurements under stated settings, and independent evaluations carry their own limitations.

Independent model evaluations

An April 2026 preprint presented an independent safety evaluation of Kimi K2.5 across chemical, biological, radiological, nuclear, and explosive information, cybersecurity, misalignment, political censorship, bias, and harmful-request handling. The authors found substantial dual-use capability and lower refusal rates than the closed models they compared on some hazardous prompts. They also found that the model did not show frontier-level autonomous vulnerability discovery and exploitation, and did not appear to have long-term malicious goals, while reporting concerning levels of sabotage ability and self-replication propensity in their tested scenarios.^[40] The work was a preliminary preprint, not a complete certification of safety or danger.

In July 2026, the United Kingdom AI Security Institute and the United States Center for AI Standards and Innovation jointly evaluated K3 on a limited set of cyber tasks. K3 performed significantly below the leading closed US models they tested but above GLM-5.2: it scored 32 percent on the ExploitBench exploit-development benchmark against 24 percent for GLM-5.2, achieved arbitrary code execution on none of the 41 tasks, and on a 32-step simulated corporate network attack reached step 17 on average, against 28.5 for the most cyber-capable US models, completing the range once in ten attempts. The evaluators reported that K3's safeguards did not prevent it from attempting exploit development or offensive cyber operations.^[41] They also warned that the comparison used only selected tasks, that the K3 aggregate estimate rested on a single 41-task benchmark, and that safeguards were disabled for the US comparison systems. Those conditions limit broader conclusions.

Distillation allegation

In February 2026, Anthropic alleged that Moonshot, DeepSeek, and MiniMax had used coordinated fraudulent accounts to extract outputs from Claude for model training. Anthropic said it attributed each campaign through IP address correlation, request metadata, and infrastructure indicators, in some cases corroborated by industry partners; for Moonshot, it said request metadata matched the public profiles of senior Moonshot staff.^[42] This is an allegation by a competitor and service provider, not a court or regulatory finding. The allegation does not by itself establish that any specific Kimi capability was copied or explain Moonshot's model performance.

Government and industry response to K3

The distillation question moved from a corporate complaint to a diplomatic one within a week of the K3 launch. On July 22, 2026, White House science and technology adviser Michael Kratsios wrote on X: "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model." He also alleged that Moonshot had acquired GB300-equipped servers and accessed GB300s in Thailand; the post did not include supporting evidence.^[88] Treasury Secretary Scott Bessent posted the same day that "When [Chinese] firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table."^[89] China's Ministry of Commerce rejected the allegations as lacking factual and legal basis and characterized the US position as "AI hegemony".^[66] The South China Morning Post reported on July 23 that Moonshot had not publicly responded to Washington's accusations, but its head of enterprise business, Huang Zhenxin, had told Chinese media earlier that week that K3 was "not a distillation or replica of existing models", in the South China Morning Post's translation, and Reuters later described Moonshot as disputing the distillation claims.^[5][90] Researchers quoted by the Post said there was insufficient public evidence to establish that K3 had been distilled from Fable 5, which Anthropic had released on June 9 but withdrawn days later under US export controls, restoring global access only on July 1.^[89][90]

The episode also produced an unusual industry intervention. On July 24, Nvidia chief executive Jensen Huang used his first post on X to publish an open letter arguing that open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. Fortune reported about 25 initial signatories including Nvidia, Microsoft, Meta, Palantir, Hugging Face, Andreessen Horowitz, Perplexity, and IBM, with OpenAI and Anthropic absent, and placed the letter in the context of the debate over restricting Chinese open-weight models such as K3; Caixin reported the count had grown to 132 organizations, including Amazon, Google, and OpenAI, by July 28.^[66][67]

Anthropic's September 2026 report

On September 10, 2026, Anthropic's threat-intelligence report "Detecting and countering misuse of AI: September 2026" went further than its February post. It alleged that Moonshot had silently forwarded some customer requests to Claude instead of processing them with Kimi and shown Claude's answers to users as if they came from Kimi. In one ten-day period, the report says, Moonshot relayed almost 300,000 customer requests, mostly to Opus, through a proxy network of 5,380 fraudulent accounts that appeared to be located mainly in Singapore and Japan.^[58] Anthropic said Moonshot saved at least a portion of these exchanges, built a pipeline to extract Claude's chain-of-thought transcripts for training, and circumvented a "thinking signature" control by replaying signatures in new sessions; it attributed more than 23 million distillation exchanges to Moonshot between May and July 2026.^[58] The report also said the relayed queries contained sensitive customer data, including requests from a user it assessed as likely affiliated with the People's Liberation Army who asked what they believed was Kimi to analyze CCTV footage of a single individual in Chengdu.^[58]

The same report made similar routing allegations against DeepSeek, said Xiaomi had also fed user conversations into Claude, and accused other Chinese developers, including Alibaba, Z.ai, and MiniMax, of distillation campaigns. Moonshot did not immediately respond to requests for comment from the South China Morning Post.^[58][91] On September 22, Gizmodo, citing The Information, reported that the Cyberspace Administration of China was investigating DeepSeek and Moonshot over the claims that they had routed user prompts and data to Claude.^[92] These are allegations by a competitor, not findings by a court or regulator.

Investor arbitration

In November 2024, GSR Ventures and four other investors in Recurrent AI, a company Yang co-founded before Moonshot, initiated arbitration at the Hong Kong International Arbitration Centre against Yang and co-founder Zhang Yutao, described in Chinese reporting as Moonshot's chief technology officer. They alleged that Yang had not completed required procedures before forming and financing Moonshot. Yang said Recurrent AI's board had approved his departure and that the necessary formalities were complete; Moonshot's lawyers said the claims lacked factual and legal basis.^[43][76][78]

A related public dispute involves a different person with a similar name. In December 2024 Allen Zhu of GSR Ventures accused Zhang Yutong, a former managing partner at his firm who joined Moonshot as a co-founder in 2024 and later became its president, of concealing a personal stake in Moonshot from Recurrent AI's other shareholders. Zhang Yutong is not a respondent in the arbitration, and Moonshot said at the time that there were no new developments in the case.^[79] These are allegations by a former investor, not findings.

The proceeding advanced procedurally and then went quiet. TechNode reported in February 2025 that both sides had paid their fees to the arbitration centre and that the tribunal had been formed without a settlement.^[77] When Huaxia Times asked about the case's progress in December 2025, Moonshot did not reply.^[76] The claims are allegations, not findings, and because Moonshot's listing application was filed confidentially, any treatment of the dispute in it was not public.^[5]

Organization and scope

Moonshot's official Chinese site lists a contact address in Beijing's Haidian district.^[1] Neither its Chinese nor its English company page states a headcount.^[1]

The registry shows who runs the company. Following the July 29, 2026 conversion to a joint stock company, Yang Zhilin is recorded as chairman and general manager, Zhou Xinyu and Zhang Yutong as directors, and Song Sijia as head of finance.^[63] Zhang Yutong, a former venture investor who joined as a co-founder, was reported to have been made president in December 2025 with responsibility for commercialization, which the Chinese business press then described as a relatively weak area for the company.^[76]

The company operates both consumer products and research releases. The Kimi service is a hosted application; Moonshot's downloadable checkpoints are model artifacts; Mooncake is serving infrastructure; and Agent Swarm is an orchestration approach. Keeping those layers separate avoids attributing a product feature, benchmark result, or license term to every Moonshot model. The 2026 releases add another layer, standalone infrastructure such as MoonEP and FlashKDA that is licensed separately from the weights it supports.^[45]

Moonshot's stated AGI ambition places it in a broader research conversation involving organizations and researchers such as Google, Yann LeCun, and Yoshua Bengio. That context does not imply a partnership or shared technical position.

See also

• Artificial intelligence
• Large language model
• AI agent
• Kimi chatbot
• Kimi Work
• Yang Zhilin
• Kimi K1.5
• Kimi K2
• Kimi K2.5
• Kimi K2.6
• Kimi K2.7 Code
• Kimi K3
• Kimi Linear
• MoonEP
• Mooncake
• DeepSeek
• Qwen
• MiniMax
• Zhipu AI
• 01.AI
• StepFun
• OpenAI
• Anthropic
• DeepMind
• Nvidia
• Samsung
• SK Hynix

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Moonshot AI. "About Moonshot AI." Accessed September 23, 2026. moonshot.ai/about ; archived version of July 30, 2026: web.archive.org/...about ; Moonshot AI. "关于我们" (About us). Accessed September 23, 2026. moonshot.cn/about
2. ^1 ^2 ^3 ^4Ingrid Lunden. "China's Moonshot AI zooms to $2.5B valuation, raising $1B for an LLM focused on long context." TechCrunch, February 21, 2024. techcrunch.com/...moonshot-ai-funding-china
3. ^Kimi Help Center. "充值与开票" (Recharge and invoicing). Accessed September 23, 2026. kimi.com/...api-billing-and-finance
4. ^1 ^2 ^3 ^4 ^5 ^6Moonshot AI. "Kimi K3." July 16, 2026. kimi.com/...kimi-k3
5. ^1 ^2 ^3 ^4 ^5Reuters, via Business Standard. "Chinese AI firm Moonshot files confidentially for Hong Kong IPO." September 3, 2026. business-standard.com/...g-kong-ipo-126090301687_1
6. ^Zhilin Yang. "Personal academic page." Accessed July 28, 2026. kimiyoung.github.io ; Carnegie Mellon University Language Technologies Institute. "Zhilin Yang." Accessed July 28, 2026. lti.cs.cmu.edu/...yang-zhilin
7. ^Zihang Dai et al. "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context." arXiv:1901.02860, 2019. arxiv.org/...1901.02860 ; Zhilin Yang et al. "XLNet: Generalized Autoregressive Pretraining for Language Understanding." arXiv:1906.08237, 2019. arxiv.org/...1906.08237
8. ^Information Technology and Innovation Foundation. "How Innovative Is China in AI?" August 2024. www2.itif.org/2024-chinese-ai-innovation.pdf
9. ^Beijing Daily. "A large-model startup founded by a post-1990 entrepreneur launches Kimi Chat." October 11, 2023. news.bjd.com.cn/...10589091.shtml ; AIbase. "Kimi Chat opens to the public and no longer requires beta qualification." November 17, 2023. news.aibase.com/...3269
10. ^1 ^2South China Morning Post. "Alibaba-backed Moonshot AI claims breakthrough with expanded Chinese-character prompt for Kimi chatbot." March 20, 2024. scmp.com/...-chinese-character-prompt-kimi-chatbot
11. ^1 ^2 ^3Ruoyu Qin et al. "Mooncake: Trading More Storage for Less Computation: A KVCache-centric Architecture for Serving LLM Chatbot." 23rd USENIX Conference on File and Storage Technologies, 2025. usenix.org/...qin
12. ^1 ^2Kimi Team et al. "Kimi k1.5: Scaling Reinforcement Learning with LLMs." arXiv:2501.12599, 2025. arxiv.org/...2501.12599
13. ^1 ^2 ^3Kimi Team et al. "Kimi K2: Open Agentic Intelligence." arXiv:2507.20534, 2025. arxiv.org/...2507.20534
14. ^1 ^2Moonshot AI. "Kimi K2 repository, model documentation, and license." Accessed July 28, 2026. github.com/...Kimi-K2 ; github.com/...LICENSE
15. ^1 ^2 ^3 ^4Moonshot AI. "Kimi K3 License." July 2026. github.com/...LICENSE
16. ^1 ^2Kimi Help Center. "Kimi overview." Accessed July 28, 2026. kimi.com/...overview ; Kimi Help Center. "What can Kimi do?" Accessed July 28, 2026. kimi.com/...capability
17. ^1 ^2Kimi Help Center. "Kimi Agent overview." Accessed July 28, 2026. kimi.com/...agent-overview ; Kimi Help Center. "Kimi Slides." Accessed July 28, 2026. kimi.com/...ppt-overview
18. ^1 ^2Kimi Help Center. "What Is Kimi Work? A Local Agent for Knowledge Workers." Accessed July 28, 2026. kimi.com/...overview
19. ^Kimi Help Center. "Kimi API overview." Accessed July 28, 2026. kimi.com/...api-overview ; Kimi Help Center. "Kimi API troubleshooting." Accessed July 28, 2026. kimi.com/...api-troubleshooting
20. ^1 ^2Moonshot AI. "Kimi K2 0905." September 5, 2025. platform.kimi.com/...kimi-k2-0905
21. ^1 ^2Moonshot AI. "Kimi K2 Thinking." November 6, 2025. platform.kimi.com/...k2-think
22. ^National Institute of Standards and Technology. "CAISI Evaluation of Kimi K2 Thinking." December 12, 2025, updated January 9, 2026. nist.gov/...caisi-evaluation-kimi-k2-thinking
23. ^Yicai Global. "Kimi K2 Thinking's Reported USD4.6 Million Training Cost Isn't Official, Moonshot CEO Says." November 2025. yicaiglobal.com/...isnt-official-moonshot-ceo-says
24. ^1 ^2Kimi Team et al. "Kimi K2.5: Visual Agentic Intelligence." arXiv:2602.02276, 2026. arxiv.org/...2602.02276
25. ^1 ^2Moonshot AI. "Kimi K2.5 repository and Modified MIT License." Accessed July 28, 2026. github.com/...Kimi-K2.5 ; github.com/...LICENSE
26. ^1 ^2Moonshot AI. "Kimi K2.6: Intelligence in Motion." April 20, 2026. kimi.com/...kimi-k2-6 ; Moonshot AI. "Kimi K2.6 model card." Accessed July 28, 2026. huggingface.co/...Kimi-K2.6
27. ^1 ^2 ^3 ^4Moonshot AI. "Kimi K3 repository and model card." Accessed July 28, 2026. github.com/...Kimi-K3 ; Moonshot AI. "Kimi K3 model repository." Accessed July 28, 2026. huggingface.co/...Kimi-K3
28. ^1 ^2 ^3Kimi Code Docs. "What's New" (changelog entries for Kimi K3, July 16, 2026; K2.8 Preview, September 11, 2026; Kimi Code Desktop and Kimi Code CLI v2.0.0, September 17, 2026). Accessed September 23, 2026. kimi.com/...whats-new
29. ^1 ^2Jingyuan Liu et al. "Muon is Scalable for LLM Training." arXiv:2502.16982, 2025. arxiv.org/...2502.16982
30. ^1 ^2Kimi Team et al. "Kimi Linear: An Expressive, Efficient Attention Architecture." arXiv:2510.26692, 2025. arxiv.org/...2510.26692
31. ^Kimi Team et al. "Kimi-VL Technical Report." arXiv:2504.07491, 2025. arxiv.org/...2504.07491
32. ^Moonshot AI. "Kimi-Dev: open coding LLM for software engineering tasks." Accessed July 28, 2026. github.com/...Kimi-Dev ; Moonshot AI. "Kimi-Dev-72B model card." Accessed July 28, 2026. huggingface.co/...Kimi-Dev-72B
33. ^Reuters, via Investing.com. "China's Moonshot AI releases open-source model to reclaim market position." July 11, 2025. investing.com/...o-reclaim-market-position-4132351 (read via the Internet Archive copy of April 25, 2026)
34. ^1 ^2 ^3Alibaba Group Holding Limited. "Annual Report on Form 20-F for the fiscal year ended March 31, 2024." May 2024. sec.gov/...baba-20240331
35. ^Bloomberg News. "Tencent Joins $300 Million Financing for China's AI Unicorn." August 5, 2024. news.bloomberglaw.com/...ing-for-chinas-ai-unicorn
36. ^South China Morning Post. "China's Moonshot AI raises US$500 million in latest funding round: report." January 1, 2026. scmp.com/...00-million-latest-funding-round-report
37. ^South China Morning Post. "Moonshot AI targets US$12 billion valuation as overseas revenue surges for Kimi models." February 18, 2026. scmp.com/...on-overseas-revenue-surges-kimi-models
38. ^1 ^2 ^3 ^4Kate Park. "China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets." TechCrunch, May 7, 2026. techcrunch.com/...nd-for-open-source-ai-skyrockets
39. ^1 ^2 ^3 ^4 ^5Reuters, via Investing.com. "China's Moonshot pauses Kimi subscriptions amid hot demand, IPO push." July 20, 2026. investing.com/...-amid-hot-demand-ipo-push-4800006 (read via the Internet Archive copy of July 20, 2026) ; Bloomberg News. "China's Moonshot in Talks on Pre-IPO Funds at $50 Billion Value." July 21, 2026. news.bloomberglaw.com/...funds-at-50-billion-value
40. ^Zheng-Xin Yong et al. "An Independent Safety Evaluation of Kimi K2.5." arXiv:2604.03121, 2026. arxiv.org/...2604.03121
41. ^UK AI Security Institute and US Center for AI Standards and Innovation. "Preliminary Assessment of Kimi K3's Cyber Capabilities." July 2026. aisi.gov.uk/...ment-of-kimi-k3s-cyber-capabilities
42. ^Anthropic. "Detecting and preventing distillation attacks." February 23, 2026. anthropic.com/...d-preventing-distillation-attacks
43. ^1 ^2 ^3 ^4South China Morning Post. "Moonshot AI founders in dispute with 5 investors in arbitration in Hong Kong." December 10, 2024. scmp.com/...pute-5-investors-arbitration-hong-kong ; TMTPost. "Moonshot AI responds to investor arbitration." December 2024. en.tmtpost.com/...7368012
44. ^National Institute of Standards and Technology. "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile." NIST AI 600-1, July 2024. doi.org/...NIST.AI.600-1
45. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8GitHub REST API. "MoonshotAI organization repository listing" (43 public repositories with creation dates, licenses, and star counts). Retrieved July 31 and September 23, 2026. api.github.com/...repos
46. ^1 ^2 ^3Moonshot AI. "MoonEP: A Perfectly Balanced Expert Parallelism Library via Dynamic Redundant Experts." GitHub repository and README, initial commit July 24, 2026. Accessed July 31, 2026. github.com/...MoonEP
47. ^1 ^2Moonshot AI. "FlashKDA: high-performance Kimi Delta Attention kernels." GitHub repository and README, created April 20, 2026. Accessed July 31, 2026. github.com/...FlashKDA
48. ^1 ^2kvcache-ai. "Mooncake: the serving platform for Kimi." GitHub repository, created June 25, 2024, Apache 2.0. Accessed July 31, 2026. github.com/...Mooncake
49. ^kvcache-ai. "AgentENV (AENV): a distributed platform for running agent environments at scale." GitHub repository, created July 23, 2026, MIT. Accessed July 31, 2026. github.com/...AgentENV
50. ^1 ^2Moonshot AI. "PerceptionBench: Evaluating Atomic Visual Perception in Multimodal Large Language Models." GitHub repository and README, accessed July 31, 2026. github.com/...PerceptionBench ; Moonshot AI. "PerceptionBench." July 24, 2026. kimi.com/...perception-bench
51. ^1 ^2Moonshot AI. "MiniTriton" README. GitHub, accessed July 31, 2026. github.com/...minitriton
52. ^1 ^2 ^3Moonshot AI. "nano-kpu" README. GitHub, accessed July 31, 2026. github.com/...nano-kpu
53. ^1 ^2 ^3 ^4 ^5Moonshot AI. "Kimi Code CLI" README. GitHub, accessed September 23, 2026. github.com/...kimi-code ; Moonshot AI. "Kimi CLI (Archived)" README. GitHub, accessed September 23, 2026. github.com/...kimi-cli
54. ^Moonshot AI. "checkpoint-engine: middleware to update model weights in LLM inference engines." GitHub repository, created September 8, 2025, MIT. Accessed July 31, 2026. github.com/...checkpoint-engine
55. ^1 ^2 ^3Moonshot AI. "moonshotai/Kimi-K2.7-Code model card." Hugging Face, accessed July 31, 2026. huggingface.co/...Kimi-K2.7-Code
56. ^MarkTechPost. "Moonshot AI Releases Kimi K2.7-Code: a Coding Model Reporting +21.8% on Kimi Code Bench v2 Over K2.6." June 12, 2026. marktechpost.com/...n-kimi-code-bench-v2-over-k2-6
57. ^1 ^2 ^3 ^4Kimi API Platform. "Models." Accessed September 23, 2026. platform.kimi.ai/...models ; Moonshot AI. "Kimi Code." Accessed July 31, 2026. kimi.com/code
58. ^1 ^2 ^3 ^4Anthropic. "Detecting and countering misuse of AI: September 2026." Threat intelligence report, September 10, 2026 (section GTG-16002). anthropic.com/...ntelligence-report-september-2026
59. ^1 ^2 ^3 ^4 ^5 ^6 ^7The Paper (Pengpai). Report on Moonshot AI's F round of more than US$3.5 billion at a US$35 billion post-money valuation, with the G round at a US$50 billion pre-money valuation. July 29, 2026. thepaper.cn/newsDetail_forward_33682828 ; HK01. Report on the same round. July 29, 2026. global.hk01.com/...60375013
60. ^1 ^2 ^3Bloomberg News, via Sina Finance. "月之暗面据悉将拆除红筹架构 以求香港IPO获批" (Moonshot AI reportedly to dismantle red-chip structure to win Hong Kong IPO approval). May 19, 2026. finance.sina.com.cn/...doc-inhymumc9940916.shtml ; TechNode. "Moonshot AI reportedly plans final pre-IPO round at $50 billion valuation." July 22, 2026. technode.com/...-ipo-round-at-50-billion-valuation
61. ^Cailian Press (Cls.cn). Report that Moonshot AI had sent a listing resolution to investors and expected a Hong Kong listing within about six months. July 19, 2026. cls.cn/...2430468
62. ^Alibaba Group Holding Limited. "Annual Report on Form 20-F for the fiscal year ended March 31, 2026," Note 4(e), Investment in Moonshot AI Ltd. Filed May 20, 2026. sec.gov/...baba-20260331
63. ^1 ^2 ^3 ^4Sina Technology. Report on Moonshot AI's conversion to a joint stock limited company and officer changes, citing Tianyancha registration data. July 30, 2026. finance.sina.com.cn/...doc-inikqqhm1884301.shtml ; MyDrivers. Report on the same registration change. July 30, 2026. news.mydrivers.com/...1140188
64. ^1 ^2Cailian Press (Cls.cn). Report on a National Social Security Fund vehicle appearing among Moonshot AI's registered shareholders and the increase in registered capital. July 10, 2026. cls.cn/...2422917
65. ^1 ^2IT Home, citing STAR Market Daily. "月之暗面 Kimi 融资获国资加持，国智投、中国移动等央企巨头入场" (Moonshot's Kimi financing gains state capital backing as Guozhitou, China Mobile and other central SOEs join). May 19, 2026. ithome.com/...336
66. ^1 ^2 ^3 ^4Caixin Global. "Moonshot Open-Sources Kimi K3 as U.S.-China AI Tensions Intensify." July 29, 2026. caixinglobal.com/...i-tensions-intensify-102468927
67. ^Fortune. "Jensen Huang just used his first ever X post to warn the AI industry not to make the mistake that software narrowly avoided in the 1980s." July 24, 2026. fortune.com/...uang-open-source-letter-nvidia-kimi
68. ^1 ^2Caixin Global. "Kimi K3 Demand Surge Forces Moonshot AI to Pause Sign-Ups." July 21, 2026. caixinglobal.com/...ai-to-pause-sign-ups-102466370
69. ^1 ^2Investing.com, via Yahoo Finance. "Cadence and Synopsys slide as Kimi K3 designs chip in 48h using no proprietary EDA." July 17, 2026. finance.yahoo.com/...opsys-slide-kimi-k3-162824133
70. ^Alina Maria Stan. "Kimi K3 spooked markets. The AI selloff was already loaded." The Next Web, July 17, 2026. thenextweb.com/...kimi-k3-china-ai-tech-rout-selloff
71. ^Kinling Lo. "China's MiniMax, Zhipu AI beat OpenAI to IPO." Rest of World, January 6, 2026. restofworld.org/...zhipu-ai-minimax-ipo
72. ^National Business Daily (Meiri Jingji Xinwen). Quarterly AI application value ranking for the first quarter of 2026, reporting Kimi monthly active users of 8.338 million and a fourth consecutive quarterly decline. April 21, 2026. nbd.com.cn/...4350446
73. ^QuestMobile. "QuestMobile 2026年AI应用市场发展半年报" (2026 AI application market half-year report). July 14, 2026. questmobile.com.cn/...2076954943839809537
74. ^Nathan Lambert. "Kimi K3: The open-weights escalation." Interconnects, July 20, 2026. interconnects.ai/...k3-the-open-weights-escalation
75. ^Hugging Face Hub API. "moonshotai" model listing with repository creation dates and download counts. Retrieved September 23, 2026. huggingface.co/...models
76. ^1 ^2 ^3 ^4Huaxia Times, via Sina Finance. Report on Zhang Yutong's appointment as president of Moonshot AI, her prior role as co-founder, and the unanswered status of the arbitration brought against Yang Zhilin and co-founder and chief technology officer Zhang Yutao. December 12, 2025. finance.sina.com.cn/...doc-inhanyzp7398096.shtml
77. ^TechNode. "Moonshot arbitration case advances amid ongoing disputes." February 25, 2025. technode.com/...ase-advances-amid-ongoing-disputes
78. ^1 ^221st Century Business Herald. Report that former Recurrent AI investors had filed arbitration at the Hong Kong International Arbitration Centre against Yang Zhilin and co-founder and chief technology officer Zhang Yutao, with Moonshot's counsel responding that the claims lacked legal and factual basis. November 11, 2024. 21jingji.com/...976e76053e24cb1ddde32fba7fdb6e59
79. ^The Paper (Pengpai). Report on Allen Zhu's allegation that Zhang Yutong concealed a Moonshot AI shareholding from Recurrent AI's other investors, and on her roles at GSR Ventures and Moonshot AI. December 5, 2024. m.thepaper.cn/newsDetail_forward_29557510
80. ^1 ^2 ^3Cailian Press (Cls.cn). Flash report, attributed only to unnamed reports, that Moonshot AI planned to submit its Hong Kong IPO application as early as August 2026 and raise about US$3 billion. August 3, 2026. cls.cn/...2444096 ; Sina Technology. Report on the same claim, including the brought-forward G round and a skeptical comment from a person close to the company. August 3, 2026. finance.sina.com.cn/...doc-inikzqsr2802606.shtml
81. ^IT Home. Report that Moonshot AI told Jiemian News the IPO filing report was untrue. August 3, 2026. ithome.com/...169 ; National Business Daily (Meiri Jingji Xinwen). Report that a person familiar with the matter called the filing report untrue. August 3, 2026. nbd.com.cn/...4530564
82. ^KrASIA, adapting IPO Zaozhidao. "Moonshot AI targets USD 50 billion valuation ahead of Hong Kong IPO." July 30, 2026. kr-asia.com/...on-valuation-ahead-of-hong-kong-ipo
83. ^1 ^2 ^3Sina Technology, via TechWeb, citing Bloomberg. Report that the F round closed ahead of schedule after heavy oversubscription and that a pre-IPO round at a US$50 billion pre-money valuation had begun. July 30, 2026. finance.sina.com.cn/...doc-inikpxkk9644803.shtml
84. ^Guandian, via Sina Finance. "月之暗面传本周秘密递表港交所 投前估值上看500亿美元" (Moonshot AI reportedly filed confidentially with HKEX this week; pre-money valuation up to US$50 billion). September 3, 2026. finance.sina.com.cn/...doc-iniqpqxy2132455.shtml
85. ^Moonshot AI. "Kimi-Audio" README and license section. GitHub, repository created April 25, 2025. Accessed September 23, 2026. github.com/...Kimi-Audio
86. ^TechCrunch. "Kimi-maker Moonshot AI targets $2B in annual revenue." September 11, 2026. techcrunch.com/...gets-2-billion-in-annual-revenue
87. ^1 ^2Shanghai Securities News, via Sina Finance. "ARR破3亿美元 月之暗面估值增至315亿美元" (ARR passes US$300 million; Moonshot AI's valuation rises to US$31.5 billion). June 30, 2026. finance.sina.com.cn/...doc-inifesxs1557054.shtml
88. ^Michael Kratsios. Post on X. July 22, 2026. x.com/...2079933645888880708
89. ^1 ^2TechCrunch. "Treasury threatens sanctions after White House claims Moonshot distilled Anthropic's Fable." July 22, 2026. techcrunch.com/...nshot-distilled-anthropics-fable
90. ^1 ^2South China Morning Post. "Global AI experts push back on US 'distillation' claims against Moonshot's Kimi K3 model." July 23, 2026. scmp.com/...claims-against-moonshots-kimi-k3-model
91. ^South China Morning Post. "Moonshot, DeepSeek secretly routed user requests to Claude, Anthropic claims." September 11, 2026. scmp.com/...-user-requests-claude-anthropic-claims
92. ^Gizmodo. "China Probes DeepSeek, Moonshot AI Over Anthropic's Claims They Route Requests to Claude." September 22, 2026. gizmodo.com/...route-requests-to-claude-2000815507
93. ^1 ^2 ^3Phoenix Finance, via Huxiu. "估值狂飙至500亿美元，月之暗面突然遭遇暗剑" (As its valuation soars to US$50 billion, Moonshot AI is suddenly hit by a hidden blade). September 12, 2026. m.huxiu.com/...4890711
94. ^Moonshot AI. "Kimi Vendor Verifier" README (K3 evaluation results by provider). GitHub, accessed September 23, 2026. github.com/...Kimi-Vendor-Verifier