Kimi and Moonshot AI: The Complete Guide (2026)

The Complete Guide to Kimi, Moonshot AI, and the Rise of Chinese LLMs

On March 19, 2026, Cursor launched Composer 2.0. They called it "frontier-level coding intelligence." The blog post did not mention where the intelligence actually came from. Within hours, a developer found the internal model identifier: accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast. The base model was Kimi K2.5, built by a 80-person Chinese startup called Moonshot AI. One of the biggest AI coding tools in the world, with $2 billion in annual recurring revenue and over a million daily active users, had quietly built its flagship product on top of a Chinese open-source model.

This guide covers everything you need to know about Kimi, Moonshot AI, and the broader Chinese LLM landscape that produced it.

This guide is written by Yuma Heymans (@yumahey), founder of o-mega.ai, the AI workforce platform where autonomous agents learn to use business tool stacks and execute workflows.

Contents

1. The Cursor Controversy: What Actually Happened
2. Who Is Moonshot AI
3. Yang Zhilin: The Researcher Behind Kimi
4. The Kimi Model Lineage
5. Kimi K2: One Trillion Parameters, Open Weights
6. Kimi K2.5: Agent Swarms and Vision
7. Benchmark Comparisons
8. API Pricing Comparison
9. Kimi's Consumer Product
10. The Chinese LLM Landscape
11. Head-to-Head: Kimi vs DeepSeek vs Qwen
12. What This Means for the Industry

1. The Cursor Controversy: What Actually Happened

The timeline is straightforward.

March 19, 2026: Cursor (owned by Anysphere, San Francisco) launched Composer 2.0 through a blog post promoting it as their most capable coding model. Moonshot AI was not mentioned anywhere in the announcement.

Within hours: A developer using the handle @fynnso on X discovered the internal model identifier string embedded in the system: accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast. The "kimi-k2p5" portion made it immediately clear that Kimi K2.5 served as the base model.

March 22, 2026: The story hit TechCrunch, VentureBeat, and eWEEK. Cursor's VP of Developer Education, Lee Robinson, acknowledged it publicly: "Yep, Composer 2 started from an open-source base! Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training."

Cursor co-founder Aman Sanger added: "It was a miss to not mention the Kimi base in our blog from the start. We'll fix that for the next model."

The license problem. Kimi K2.5 uses a Modified MIT License with a specific clause: if the derivative work is used in a commercial product exceeding 100 million monthly active users or generating more than $20 million per month in revenue, the licensee must "prominently display 'Kimi K2.5' on the user interface." Anysphere's $2 billion ARR translates to roughly $166 million per month, more than 8x the threshold. Moonshot AI stated that Cursor used the model "as part of an authorized commercial partnership" through Fireworks AI, suggesting a separate commercial arrangement existed. But the lack of upfront disclosure remained the central issue.

Community reaction was mixed. Some praised the quality of Chinese open-source models. Others questioned whether Cursor was primarily a "model routing layer" with a good UI rather than an independent AI research company. The political dimension was unavoidable: an American AI darling had built its core product on Chinese technology during a period of intense U.S.-China AI competition rhetoric.

The incident did more for Kimi's brand recognition outside China than any marketing campaign could have. It also raised fundamental questions about attribution, licensing compliance, and the actual origin of AI capabilities in commercial products.

2. Who Is Moonshot AI

Company name: Moonshot AI (Chinese: 月之暗面, literally "The Dark Side of the Moon," a tribute to Pink Floyd that reflects CEO Yang Zhilin's love of classic rock)

Founded: March 2023, Beijing, China

Founders: Yang Zhilin, Zhou Xinyu, and Wu Yuxin. All three are Tsinghua University alumni and former bandmates in a group called "Splay."

Mission: Build foundation models to achieve AGI.

Headcount: Approximately 80 employees.

Product: Kimi, both a consumer chatbot and an API platform.

Funding History

Total raised: Over $1.77 billion across multiple rounds from 8+ investors.

Moonshot AI became the fastest Chinese company to reach decacorn status (valuation exceeding $10 billion), achieving the milestone in roughly two years. ByteDance took over four years. Pinduoduo took over three.

Revenue: $240 million reported by November 2025. After the Kimi K2.5 launch, the company reported that cumulative revenue in fewer than 20 days already exceeded the entire 2025 annual total.

An 80-person company generating hundreds of millions in revenue and valued at $18 billion. Those are exceptional numbers for any AI company, let alone one operating from Beijing.

3. Yang Zhilin: The Researcher Behind Kimi

Yang Zhilin is one of the most technically credentialed founders in the AI industry, on either side of the Pacific.

Born: 1992 in China.

Education: BSc in Computer Science at Tsinghua University, graduated top of his class in 2015. He had no programming background before university. He won first prize at the National Olympiad in Informatics in Guangdong province, which earned him admission to Tsinghua.

PhD: Carnegie Mellon University, completed in 2019 after only four years (the standard timeline is six). His advisors were Ruslan Salakhutdinov (who later became Apple's head of AI research) and William Cohen (a principal scientist at Google DeepMind).

Research contributions during his PhD: He interned at Google Brain and Meta's FAIR (Facebook AI Research). He co-developed two architectures that shaped the trajectory of language modeling:

• Transformer-XL: Extended the standard Transformer to handle longer sequences through a recurrence mechanism. This was foundational work for long-context language models.
• XLNet: A language model that surpassed Google BERT on 20 standard NLP tasks and achieved state-of-the-art results on 18. Selected as an Oral presentation at NeurIPS 2019, one of the highest distinctions at the conference.

After completing his PhD, Yang returned to China and co-founded Moonshot AI in March 2023. His deep expertise in long-context modeling directly influenced Kimi's initial differentiator: supporting 200,000 Chinese characters in a single conversation when the model first launched, a world record at the time.

Co-founders:

• Zhou Xinyu: Deep engineering and systems-level expertise. Responsible for moving models efficiently from research to production at scale.
• Wu Yuxin: Research background in vision, multimodal modeling, and open-source culture. Previously at Google Brain and Facebook AI Research.

4. The Kimi Model Lineage

Kimi's development shows a clear trajectory from consumer chatbot to open-weight frontier model.

Kimi v1.0 (October 2023)

The initial release supported up to 200,000 Chinese characters per conversation. At the time, this was the longest context window of any publicly available LLM. By early 2024, Kimi had expanded this to over 2 million characters in a single conversation. This long-context capability was directly connected to Yang Zhilin's research on Transformer-XL during his PhD.

Kimi k1.5 (January 20, 2025)

The first major architectural leap. Key characteristics:

• Training approach: Reinforcement learning (RL) with long chain-of-thought reasoning
• Context window: 128,000 tokens (RL context scaled to 128K)
• Multimodal: Vision and text capabilities
• Training innovations: Online mirror descent for policy optimization, partial rollout reuse for efficiency
• What it did NOT use: Monte Carlo tree search, value functions, or process reward models

The k1.5 paper (arXiv 2501.12599) demonstrated that carefully scaled reinforcement learning alone, without elaborate search techniques, could match OpenAI's reasoning models on core benchmarks.

Kimi K2 (July 11, 2025)

The open-weight frontier model that put Moonshot on the global map. More details in Section 5.

Kimi K2 Thinking (November 6, 2025)

A reasoning-focused variant of K2 with extended thinking capabilities. This model was the first to demonstrate stable tool use across 200-300 sequential calls. More details in Section 6.

Kimi K2.5 (January 27, 2026)

The multimodal, agent-swarm capable model that Cursor used as the base for Composer 2.0. More details in Section 6.

Kimi K3 (July 2026)

The current flagship, and the model every earlier entry in this list now redirects to. K3 is a 2.8-trillion-parameter Mixture of Experts that activates 104 billion parameters per token, selecting 16 experts from a pool of 896 across 93 layers, and it replaces K2's attention stack with 69 Kimi Delta Attention layers plus 24 Gated MLA layers. Context is 1,048,576 tokens, thinking mode is always on, and vision is native. Moonshot describes it as the first open-source model in the 3-trillion-parameter class. The older models are gone from the API rather than merely deprioritised: Moonshot discontinued the kimi-k2 series on May 25, 2026 and kimi-k2.5 on August 31, 2026, and the served lineup is now kimi-k3, kimi-k2.7-code and kimi-k2.6.

5. Kimi K2: One Trillion Parameters, Open Weights

Kimi K2, released July 11, 2025, marked Moonshot AI's arrival as a serious contender in the global foundation model competition. The numbers tell the story.

Architecture

K2 uses a Mixture of Experts (MoE) architecture with 384 experts but only activates 32 billion parameters for any given input. This means it delivers performance comparable to much larger dense models while requiring compute equivalent to running a 32B parameter model.

The MuonClip Optimizer

One of K2's most significant technical contributions is MuonClip, a novel optimizer that integrates the Muon algorithm with a QK-Clip stability mechanism. The result: K2 trained on the full 15.5 trillion tokens without a single loss spike. Training instability is one of the most expensive problems in LLM development, often requiring checkpoint rollbacks that waste days of GPU time. Solving this problem at the trillion-parameter scale is a genuine engineering achievement.

License

K2 uses a Modified MIT License. The model weights are open and available on Hugging Face under moonshotai/Kimi-K2-Instruct and moonshotai/Kimi-K2-Base. Commercial use is permitted with one condition: if the product exceeds $20M/month in revenue or 100M monthly active users, the Kimi K2 branding must be prominently displayed. This is the same license clause that created the Cursor controversy.

6. Kimi K2.5: Agent Swarms and Vision

Released January 27, 2026, K2.5 extends K2 with two major capabilities: native multimodal vision and parallel agent swarm execution.

MoonViT-3D Vision Encoder

K2.5 introduces MoonViT-3D, a 400-million parameter vision encoder based on SigLIP-SO-400M. It uses the NaViT packing strategy for variable-resolution images, supporting:

• Images: Up to 4K resolution (4096x2160) in PNG, JPEG, WebP, GIF formats
• Video: Up to 2K resolution (2048x1080) in MP4, MPEG, MOV, AVI, FLV, WebM formats
• Documents: PDF and text

The total parameter count with the vision encoder reaches approximately 1.04 trillion.

Agent Swarm: Parallel Multi-Agent Execution

This is K2.5's most distinctive capability. Trained with Parallel-Agent Reinforcement Learning (PARL), the model can self-direct up to 100 sub-agents executing up to 1,500 coordinated steps in parallel. On wide-search tasks (competitive analysis across 50+ websites, multi-file codebase debugging), the agent swarm delivers 4.5x faster execution compared to single-agent models.

Kimi K2 Thinking (November 6, 2025)

The reasoning variant deserves its own mention. Key specifications:

• Parameters: 1 trillion total, 32 billion active
• Context window: 256K tokens
• Quantization: Native INT4
• Agentic reasoning: Stable tool use across 200-300 sequential calls

K2 Thinking was evaluated by NIST's CAISI (Center for AI Safety and Intelligence) in December 2025, ranking as the #1 open-source model on multiple benchmarks at the time of its release.

7. Benchmark Comparisons

The tables in this section capture the field as it stood at K2.5's launch in January 2026, measured against the frontier models of that moment. Kimi K3's own numbers follow at the end of the section.

Core Reasoning and Math

Coding

Agentic and Tool Use

The HLE (Humanity's Last Exam) scores are particularly notable. K2.5 outperformed both Claude Opus 4.5 and GPT-5.2 by substantial margins when tools were available. The "+20.1pp" gain from adding tools (compared to without tools) suggests K2.5 was specifically optimized for agentic tool-use scenarios.

Vision and Multimodal

K2.5's vision benchmarks are strong across the board. The VideoMMMU score of 86.6% reflects the MoonViT-3D encoder's ability to process video content effectively.

Key Takeaways from Benchmarks

1. On pure math and reasoning, Kimi K2 Thinking and K2.5 matched or beat the frontier of their moment, set by GPT-5.2 and Claude Opus 4.5.
2. On coding (SWE-Bench), Kimi trailed Claude and GPT but significantly outperformed DeepSeek V3 and was competitive with Qwen3 Coder.
3. On agentic tasks with tools, K2.5 led. Its 50.2% HLE score with tools was the highest published result at the time.
4. On vision tasks, K2.5 led on MMMU Pro and MathVision, strong performance for an open-weight model.

Kimi K3 has since moved all of those numbers. Moonshot's model card reports 93.5 on GPQA Diamond, 91.2 on BrowseComp, 88.3 on Terminal-Bench 2.1, 90.0 on Video-MME, 81.6 on MMMU-Pro and 67.5 on DeepSWE. The suites moved too, which is the more useful observation. The benchmarks Moonshot chose to publish for K3 are agentic and long-horizon (terminal sessions, open-web browsing, software engineering over a real repository) rather than the single-turn math sets that dominated the K2.5 card. Saturation pushed the whole industry off MATH-500 and AIME inside a year.

8. API Pricing Comparison

One of the most compelling aspects of Chinese LLMs is pricing. The cost differences versus Western models are real, though at the very top of each vendor's range they have narrowed sharply.

Per-Million-Token Pricing

DeepSeek rates are off-peak. Its peak window (01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday) is billed at exactly double. The Claude and GPT figures are standard first-party list prices from Anthropic and OpenAI.

Kimi's Pricing (by model)

The context-length tiers are gone. Moonshot charges one flat rate across the entire 1M-token window on K3, so a 900,000-token request costs the same per token as a 9,000-token one. Cache hits are billed at 10% of the fresh input rate, a 90% saving on repeated content, which keeps Kimi cost-effective for applications with consistent system prompts or reference documents.

Cost at Scale: 100 Million Tokens Per Month

(Figures assume 100M input tokens plus 100M output tokens, with DeepSeek billed at off-peak rates.)

The interesting number here is the one that disappeared. A year ago the Chinese flagship undercut the American flagship by 5x to 10x. Today Kimi K3 costs roughly $1,800 for that workload against $3,000 for Claude Opus 5, a 1.7x gap, and Claude Sonnet 5 at $2.00 in and $10.00 out is cheaper than K3 outright. Moonshot priced a frontier model like a frontier model.

The discount did not vanish, it moved down a tier. DeepSeek V4 Flash serves a 1M-token context for $0.22 per million input tokens, roughly 45x cheaper than GPT-6 Astra, and Kimi K2.6 sits at $0.95. For startups and enterprises the question is no longer "Chinese model or Western model" but which rung of a four-rung price ladder a given task actually needs.

9. Kimi's Consumer Product

Kimi is not just an API. It operates a consumer chatbot that competes directly with ChatGPT and other conversational AI products.

Platform: Available at kimi.com and kimi.ai, plus mobile apps.

Peak usage: 36 million monthly active users in China, reaching 50 million globally by August 2025.

Market position in China: Kimi ranked among the top Chinese AI chatbots but dropped to 7th in active monthly users by June 2025 as competition from Baidu's Ernie, Alibaba's Tongyi Qianwen, and others intensified.

International expansion: After the K2.5 launch, overseas revenue overtook domestic income. International subscriber growth accelerated significantly, and overseas API revenue quadrupled since November 2025.

Key differentiator: Kimi's original claim to fame was its long-context capability. The ability to process entire books, lengthy documents, and extended conversations in a single session gave it a distinct positioning in the Chinese market. When K2 launched with open weights, this shifted the conversation toward Kimi as a platform for developers and enterprises rather than just a consumer chatbot.

10. The Chinese LLM Landscape

Kimi does not exist in isolation. It emerged from a Chinese AI ecosystem that has produced several world-class foundation models. Understanding the broader landscape provides context for why Chinese LLMs have advanced so rapidly.

DeepSeek (High Flyer Capital Management)

Background: DeepSeek is a research lab funded by High Flyer, a quantitative hedge fund based in Hangzhou. The hedge fund's computing resources gave DeepSeek access to GPU clusters that most Chinese startups could not afford.

Key models:

DeepSeek's defining characteristic is efficiency. Training V3 for only $5.6 million (2.79 million GPU-hours) while matching or exceeding models that cost hundreds of millions to train was a landmark moment for the industry. It demonstrated that brute-force compute spending is not the only path to competitive performance.

Pricing: DeepSeek has stayed the price floor. The current lineup runs $0.22 in and $0.66 out per million tokens for V4 Flash, and $0.66 in and $1.98 out for V4 Pro at off-peak rates, both on a 1M-token context, and a cache hit on V4 Flash costs $0.007 per million tokens. Peak-hour traffic is billed at double, a scheduling incentive no Western provider offers.

Open source: Yes, fully open weights under permissive licenses.

Qwen (Alibaba Cloud)

Background: Alibaba's LLM family, developed by the Tongyi Lab division of Alibaba Cloud.

Key models:

Market share: 32.1% of China's enterprise LLM market in H2 2025, nearly doubling from 17.7% in H1 2025.

Noteworthy: Alibaba is simultaneously Moonshot AI's largest investor and its direct competitor. Alibaba led Moonshot's $1B Series B extension in February 2024 and has participated in subsequent rounds. This kind of "invest in your competitors" dynamic is common in the Chinese tech ecosystem, where major platforms often back multiple horses in emerging categories.

Yi / 01.AI (Kai-Fu Lee)

Founder: Kai-Fu Lee, former president of Google China and former executive at Microsoft and Apple.

Founded: March 2023 (same month as Moonshot AI).

Key model: Yi-34B achieved MMLU 76.3%, outperforming Meta's Llama 2-70B (68.9%) and TII's Falcon 180B despite being a much smaller model.

Current status: 01.AI remains active but has not released a model at the frontier scale of Kimi K2 or DeepSeek V3. The company focuses on efficient, smaller-scale models for enterprise deployment.

Ernie (Baidu)

Model: Ernie Bot 4.0

Baidu's LLM benefits from deep integration with China's dominant search engine. Ernie consistently leads on Chinese language tasks and has gradually closed the gap with GPT-4 on general benchmarks. Its primary moat is the Baidu ecosystem: search, maps, cloud, and enterprise services that give Ernie distribution advantages within China.

GLM / ChatGLM (Zhipu AI)

Model: GLM-4.5 with 355 billion parameters, GLM-4.5-Air at 106 billion.

Zhipu AI has focused on the Chinese enterprise market with strong multimodal capabilities. GLM-4.5 approaches GPT-4 level performance on general benchmarks. Open-source versions are available.

Overview: Chinese LLM Ecosystem

11. Head-to-Head: Kimi vs DeepSeek vs Qwen

These three families represent the current frontier of Chinese AI, and all three turned over in 2026: Kimi K3 in July, DeepSeek V4 over the summer, and Qwen3.8-Max, announced August 3, 2026. Each has distinct strengths.

Architecture Comparison

Benchmark Comparison

The three labs no longer publish the same suites, which itself says something about where the competition moved. These are each lab's own headline numbers for its current flagship:

Self-reported numbers on non-overlapping benchmarks are not a ranking. Read them as each lab's statement of what it optimised for: Moonshot for long-horizon agentic work in a terminal and a browser, DeepSeek for raw coding and competitive programming, Alibaba for human preference in production.

When to Use Which

The Licensing Factor

This is where the models differ significantly:

For large-scale commercial applications, DeepSeek and Qwen have simpler licensing stories. Moonshot carried the branding requirement forward into K3 unchanged and added a second clause on top of it: reselling the weights as a hosted service above $20 million of revenue over any consecutive twelve months now requires a negotiated agreement with Moonshot before deployment, not just a logo in the interface. That is the Cursor lesson written directly into the license.

12. What This Means for the Industry

The Kimi story illustrates several trends that will define AI development in 2026 and beyond.

Chinese open-source models are at the frontier. This is no longer debatable. Kimi K3 shipped 2.8 trillion open weights in the same quarter the leading American labs shipped closed ones. DeepSeek V4 Pro reports 80.6 on SWE-Bench Verified with an MIT license attached. Qwen3.8-Max sits inside the LMArena top five. The gap between Chinese and American models has narrowed to the point where leadership shifts depending on which benchmark you examine.

Open weights change the competitive dynamics. Cursor built Composer 2.0 on Kimi K2.5 because the model was available, capable, and cheap. This is exactly how open-source software has always worked. The difference is that we are now seeing it with models that cost tens of millions of dollars to train, being fine-tuned and deployed by companies that would otherwise need to invest hundreds of millions in their own training runs.

The "model routing layer" is a viable business. Cursor's $2 billion ARR demonstrates that you do not need to train your own foundation model to build a massive AI business. What matters is the product experience, the fine-tuning for specific use cases, and the integration quality. This pattern will repeat across every vertical.

Pricing pressure is structural, but it stopped being a China discount. DeepSeek V4 Flash still serves a 1M-token context at $0.22 per million input tokens against $10.00 for GPT-6 Astra. Kimi K3, however, lists at $3.00 in and $15.00 out, above Claude Sonnet 5 at $2.00 and $10.00. The pressure now runs vertically inside each vendor's own catalogue rather than horizontally across the Pacific. Developers and enterprises will increasingly adopt multi-model strategies, using the cheapest model that meets their quality requirements for each specific task.

The 80-person company is the new model. Moonshot AI built a $18 billion company with 80 people. DeepSeek trained a frontier model for $5.6 million. These are not anomalies. They reflect a structural shift in how AI companies operate. The capital intensity of AI training is real, but the human capital requirements are lower than the industry assumed.

Agent capabilities are the new frontier. Kimi K2.5's agent swarm, with 100 sub-agents executing 1,500 coordinated steps, pointed to where the competition was heading, and K3 confirms it. Raw benchmark scores on math and coding are approaching saturation, which is exactly why Moonshot's K3 card leads with Terminal-Bench and BrowseComp instead of MATH-500. The next differentiation vector is how well models orchestrate complex, multi-step tasks autonomously, across a million tokens of context and hundreds of tool calls, without a human resetting the loop.

The rise of Kimi and Chinese LLMs broadly is not a threat narrative. It is an efficiency narrative. These models demonstrate that world-class AI can be built with smaller teams, lower budgets, and open distribution. The entire industry benefits when the cost of intelligence decreases. The question is not whether Chinese models will be competitive. They already are. The question is how the ecosystem adapts to a world where frontier AI capabilities are available to anyone who can download weights from Hugging Face.