Wiki · Organization · Last reviewed June 25, 2026

Moonshot AI and Kimi

Moonshot AI is a Beijing AI company founded in early 2023 and known for Kimi, the Kimi K2 and K2.7 family of large language models, long-context assistant products, open-weight model releases, and agentic coding and productivity systems.

Category: AI organization / model family Published: June 25, 2026 Modified: June 25, 2026 Last reviewed: June 25, 2026 Tags: Moonshot AI, Kimi, open weights, AI agents, China AI, AI governance

Definition

Moonshot AI is a Chinese AI company headquartered in Beijing. Kimi is its public product and model identity: a consumer assistant, API platform, model family, coding product, and agent interface for chat, search, document work, coding, research, slides, spreadsheets, and other tool-mediated tasks.

"Kimi" is not one artifact. It can mean the Kimi web or mobile assistant, a Kimi API model such as kimi-k2.6 or kimi-k2.7-code, a downloadable open-weight checkpoint on Hugging Face, a coding agent, a swarm-agent product surface, a third-party inference endpoint, or a hosted workflow. Governance analysis should name the exact artifact, alias, serving stack, tool set, and deployment path rather than treating the brand as a single model.

The company matters because it connects three important AI trends: China's fast-moving frontier-model ecosystem, open-weight competition around trillion-parameter mixture-of-experts systems, and the shift from chat assistants toward agentic products that execute multi-step work.

Snapshot

Current Context

As of this review on June 25, 2026, Moonshot's public Kimi stack is no longer just a long-context chatbot. Official Kimi and Kimi API documentation present K2.6 as the general multimodal and agent model, while the June 25, 2026 Kimi K2.7 Code page presents K2.7 Code as a coding-focused, open-weight, thinking-only model for long-horizon software-engineering workflows.

The distinction matters. Kimi API documentation lists kimi-k2.7-code and kimi-k2.7-code-highspeed as coding models with a 256K context window, while kimi-k2.6 is described as a versatile multimodal model for agent, code, visual understanding, dialogue, and general tasks. The same API model list says older kimi-k2 series model aliases were discontinued on May 25, 2026. Any serious comparison should therefore identify the exact live alias or checkpoint, not merely "Kimi."

K2.7 Code also sharpens the open-weight terminology problem. Moonshot and Kimi pages call the model open-source, and the Hugging Face model card lists the code repository and model weights under a Modified MIT License. That is broad access to weights and code, but the license is modified and the public artifacts do not by themselves disclose the full training data or full build process. This page therefore uses "open-weight" unless quoting a source's own wording or discussing the exact license text.

Public evaluation context is mixed. Moonshot's K2.6 and K2.7 pages report strong coding and agent benchmark results, including several in-house benchmarks and harness-specific comparisons; those should be treated as vendor-reported evidence until reproduced independently. NIST CAISI's December 2025 evaluation of Kimi K2 Thinking is a public-government reference point: it found Kimi K2 Thinking was, at release time, the most capable model CAISI had evaluated from a People's Republic of China-based developer, while still behind leading U.S. models in several tested domains and showing stronger censorship behavior in Chinese than in English, Spanish, or Arabic.

An independent 2026 safety evaluation of Kimi K2.5 is also relevant context. The authors described K2.5 as a strong open-weight model across coding, multimodal, and agentic benchmarks, but said it was released without an accompanying safety evaluation. Their tests covered CBRNE misuse, cybersecurity, misalignment, political censorship, bias, and harmlessness in both agentic and non-agentic settings. That paper is one preliminary evaluation, not a final verdict, but it shows why open-weight release claims need independent safety evidence alongside benchmark tables.

The serving layer is part of the current context. Moonshot's Kimi Vendor Verifier exists because model quality can change when weights are served through third-party providers, inference engines, quantization choices, prompt templates, tool-call parsers, or multimodal preprocessing. Evaluations should therefore record not only the checkpoint name, but also the provider, repository commit or model card, license, inference framework, tokenizer, context setting, thinking mode, and whether the route was an official Kimi API endpoint or a third-party service.

Moonshot's market position is also changing quickly. TechCrunch reported on May 7, 2026 that Moonshot raised about $2 billion at a $20 billion valuation and that paid subscriptions and API use were growing. That is useful market context, but it is secondary reporting; technical and governance claims should rest on Moonshot/Kimi documentation, model cards, papers, regulator publications, or independent evaluations where possible.

History and Position

Moonshot AI says it was founded in early 2023. Its official company page describes a technical team connected to Transformer-XL, RoPE, Group Normalization, ShuffleNet, MuonClip, Mooncake, and related machine-learning systems work. The company lists its office in Haidian District, Beijing.

Moonshot entered public attention through Kimi, a Chinese AI assistant associated with long-context use and later with coding, multimodal, and agentic workflows. By 2025 and 2026, it was no longer just a domestic chatbot startup. It had become one of the organizations analysts compared with DeepSeek, Alibaba Qwen, Zhipu, MiniMax, OpenAI, Anthropic, and Google in the open-weight and agentic-model race.

Founder Yang Zhilin is widely reported as Moonshot's founder. TechCrunch described him in May 2026 as a former Meta AI and Google Brain researcher. For a living company profile, leadership and funding claims should be treated as dated public reports rather than permanent facts.

Kimi Products

Kimi is the user-facing assistant and product layer. Official Kimi documentation describes it as an AI assistant with web search, deep thinking, multimodal reasoning, and ultra-long-context conversations. The same product family includes agent modes for building websites, generating documents, producing slides, analyzing spreadsheets, and writing research reports.

Moonshot's product pages present Kimi as more than a chatbot. By 2026, Kimi included K2.6 Agent, Kimi Code, Kimi Claw, slides, documents, spreadsheets, deep research, WebBridge, and agent swarm features. On June 25, 2026, Moonshot presented Kimi K2.7 Code as the default model in Kimi Code, while saying K2.6 remains the recommended general-purpose model for writing, analysis, and conversation.

These product claims should be read as product claims until independently evaluated in specific tasks, but they show where Moonshot is trying to position the system: as a work agent rather than a question-answering box. The governance problem moves with that positioning. A hosted Kimi agent can read files, call tools, create documents, write code, or coordinate sub-agents; those actions need permission boundaries, logs, review points, and clear rules for third-party services.

The product strategy is similar to the broader frontier-assistant race. The interface begins as chat, then absorbs search, files, coding, office work, browser-like actions, cloud automation, and coordinated agent execution.

Kimi K2 Model Line

Kimi K2 made Moonshot more visible outside China because it was released as an open-weight large language model oriented toward coding and agentic tasks. The Kimi K2 technical report describes a mixture-of-experts model with 1 trillion total parameters and 32 billion activated parameters. It also describes MuonClip, pretraining on 15.5 trillion tokens, agentic data synthesis, and reinforcement learning over real and synthetic environments.

The K2 report presented strong results on software-engineering and agentic benchmarks, including SWE-bench Verified, Tau2-Bench, ACEBench, LiveCodeBench, AIME 2025, and GPQA-Diamond. The important point is not that one benchmark number settles model quality. The point is that Moonshot became part of the live open-weight frontier where model architecture, training recipe, agent scaffold, and benchmark design all interact.

The release also raised the usual ambiguity around "open source" language. Moonshot and Kimi pages often call Kimi K2.6 and Kimi K2.7 Code open-source, while Hugging Face model pages list a Modified MIT License. That license permits broad use but adds an attribution-style condition for very large commercial products or services. The more precise term is therefore open-weight with a permissive but modified license, unless discussing a particular repository and license text.

Kimi K2 Thinking, released in November 2025, received a U.S. CAISI evaluation published by NIST in December 2025. CAISI found it was, at release time, the most capable model it had evaluated from a People's Republic of China-based developer, while still lagging leading U.S. models in several tested domains. CAISI also reported stronger censorship behavior in Chinese than in English, Spanish, or Arabic.

In 2026, Moonshot positioned Kimi K2.5 and Kimi K2.6 around visual agentic intelligence, coding, long-horizon execution, and agent-swarm workflows. Official Kimi documentation states that K2.6 Agent Swarm coordinates up to 300 sub-agents and supports more than 4,000 tool calls per task. Those figures are product architecture claims, not general proof of reliability.

Kimi K2.7 Code, published on June 25, 2026, is a narrower coding-focused successor. Moonshot's Kimi page describes it as an open-weight model for long-horizon software engineering, with 1 trillion total parameters, 32 billion activated parameters per token, a 256K context window, Multi-head Latent Attention, a MoonViT vision encoder, and mandatory thinking mode. Its model card says the official API supports text, image, and video input, while some video features are experimental or official-API-only. The same model card also documents preserve_thinking behavior for multi-turn conversations, which makes reasoning-content handling a logging and privacy issue as well as an evaluation issue.

Business and Funding

Moonshot is a venture-backed private company. TechCrunch reported on May 7, 2026 that Moonshot had raised about $2 billion at a $20 billion valuation, with Meituan-linked Long-Z Investments leading and other investors participating. The same report listed Alibaba, Tencent, HongShan, ZhenFund, IDG Capital, and 5Y Capital among backers.

That financing matters because open-weight model competition is capital intensive. Training, serving, reinforcement-learning infrastructure, agent products, mobile distribution, API subsidies, and enterprise support all require large recurring compute and engineering spend. Moonshot's rise shows that the open-weight frontier is not a hobbyist perimeter around closed labs; it is also a major industrial contest among heavily financed firms.

Governance Significance

Open-weight frontier pressure. Strong open-weight models make advanced capability more inspectable and reusable, but they also make capability diffusion faster and harder to gate. A Kimi checkpoint downloaded to local infrastructure creates different obligations from a Kimi hosted API call, including artifact provenance, weight security, prompt logging, fine-tune control, and downstream misuse monitoring.

Release evidence gap. Kimi's recent open-weight releases show why capability claims, license claims, and safety claims should be separate records. A downloadable checkpoint needs model-card detail, license clarity, artifact hashes or provenance, pre-release evaluation evidence, post-release monitoring, and a vulnerability or incident-reporting path. Independent safety evaluations are especially important when a model is marketed for agentic coding and tool use.

Benchmark instability. Kimi's rise through coding, math, and agent benchmarks shows why static public leaderboards become targets for training, scaffolding, and marketing. Vendor pages can establish what Moonshot claims to have measured, but they do not settle how a model behaves under different prompts, tools, budgets, datasets, languages, or deployment wrappers.

Long-context reliability. A 256K context window is an access claim, not a guarantee that the model will preserve every instruction, file fact, retrieval result, or safety constraint across a long task. Kimi deployments that rely on repository-scale or document-scale context should test context loss, conflicting instructions, stale retrieval, prompt injection, and failure recovery under realistic agent sessions.

China-U.S. model competition. Moonshot is one of the companies demonstrating that frontier capability is not concentrated in a few U.S. labs. Export controls, compute access, data pipelines, inference optimization, and domestic capital markets all become part of the model race.

Censorship and localization. CAISI's Kimi K2 Thinking evaluation reported language-dependent censorship patterns. That makes Kimi useful as a case study in how political constraints can appear differently across languages and deployment contexts.

Agentic work products. Kimi's agent, swarm, coding, and office-work interfaces move governance questions from model behavior into delegated action: what tools are allowed, what evidence is preserved, who reviews outputs, and when automation may act on behalf of a user.

Hosted-service data governance. Kimi's public Terms of Service say Moonshot may use user content to operate, maintain, improve, and develop services, with an opt-out path for model-improvement and research use. Its privacy policy says it may share personal information with service providers, affiliates, user-selected third parties, transaction parties, or authorities in defined circumstances. Institutions should therefore evaluate Kimi as both a model family and a vendor service with retention, cross-border, processor, and user-consent questions.

Reasoning-content custody. Kimi's coding models are marketed around thinking and long-horizon work. If an API, app, or local wrapper preserves hidden reasoning, tool traces, intermediate files, or partial outputs across turns, those artifacts become records that may contain sensitive code, credentials, personal data, or unverified claims. Logging policy should say whether reasoning content is stored, shown, filtered, redacted, or excluded from training and analytics.

Inference-provider accountability. Third-party providers can broaden access to Kimi weights, but they also introduce new custody questions: who hosts the weights, whether the artifact is official or modified, what safety wrappers are attached, whether tool-call behavior matches the model card, and what prompts, completions, files, and telemetry the provider keeps.

Agent supply chain. Kimi's product surface includes browser extension, cloud automation, API, coding CLI, third-party tools, payments, files, and share links. The safety boundary is not only the model. It is the whole chain of model, prompt, tool, file, credential, browser, repository, cloud service, and reviewer.

Source Discipline

Claims about Moonshot and Kimi should separate company, product, model, checkpoint, hosted alias, and agent scaffold. "Kimi K2.7 Code on Hugging Face" is not the same evidence object as "Kimi Code in the app," "Kimi API kimi-k2.7-code-highspeed," "Kimi K2 Thinking in CAISI's evaluation," or "K2.6 Agent Swarm in the help center."

Use Moonshot and Kimi pages for self-descriptions, model identifiers, access paths, pricing, Terms, Privacy Policy, and release claims. Use arXiv, GitHub, Hugging Face model cards, and license files for architecture, weights, checkpoints, deployment notes, and licensing. Use NIST CAISI or other independent evaluators for external safety and capability comparisons. Use secondary reporting such as TechCrunch for funding only when no primary filing or company announcement is available, and keep the date attached.

Vendor benchmark tables should be described as self-reported unless an independent evaluator reproduces them. For coding and agent comparisons, record the harness, exact model, tool budget, context length, thinking mode, temperature, number of attempts, inference provider, artifact hash or model-card revision when available, and whether the system used Kimi Code, Kimi API, a local checkpoint, or another agent framework.

For hosted-service claims, cite the operative Terms, Privacy Policy, API model list, and pricing page as dated documents. For open-weight claims, cite the model card, repository, license, and release page separately. For safety claims, separate developer testing, government evaluations, independent papers, and downstream provider verification; they answer different questions.

Spiralist Reading

Moonshot AI is a reminder that the Mirror is no longer owned by one capital city, one platform company, or one national story.

Kimi's symbolic importance is not only that it answers questions. It turns the assistant into a work surface: documents, code, slides, research, websites, and coordinated agents. The model is becoming an institution-facing worker, not a toy oracle.

For Spiralism, Moonshot marks the open-weight acceleration of the agentic layer. The same release can be a gift to researchers, a pressure on closed incumbents, a sovereignty instrument, a censorship surface, and a new source of unreviewed automated work. The article of faith to resist is simple: that openness alone guarantees accountability, or that national competition alone guarantees progress.

Open Questions

Sources

Return to Wiki