Kimi: How Moonshot AI's Long-Context Model Challenges GPT and Claude

🌏 中文版

In 2023, while the rest of the world was scrambling to catch up with OpenAI, a Chinese startup called Moonshot AI made a different bet — instead of rushing to build “China’s ChatGPT,” they wagered everything on a technical direction that few were paying attention to at the time: ultra-long context windows.

That choice made Kimi one of the most recognizable AI products in the Chinese market. By 2026, it’s no longer just a “long-text tool” — it’s one of the few open-source models in the global AI race that can simultaneously challenge GPT, Claude, and Gemini across multiple dimensions.

Who Is Moonshot AI?

Moonshot AI was founded by Yang Zhilin in March 2023. The company’s Chinese name literally translates to “The Dark Side of the Moon,” taken from Pink Floyd’s classic album The Dark Side of the Moon, and it was established on the album’s 50th anniversary. Born in 1993, Yang Zhilin graduated from Tsinghua University and went on to pursue his PhD at Carnegie Mellon University, where he contributed to research on Transformer-XL and XLNet — two projects that directly advanced long-sequence modeling techniques.

Co-founders include Zhou Xinyu and Wu Yuxin, both Tsinghua alumni.

Yang Zhilin once laid out three milestones on the path to AGI: (1) long context length, (2) multimodal world models, and (3) a general architecture capable of continuous self-improvement. Kimi’s product evolution has essentially been tackling these three objectives in sequence.

Funding History

Moonshot AI is one of the fastest-funded AI companies in China, with cumulative funding exceeding $1.7 billion:

Kimi’s Starting Point: Ultra-Long Context Window

From 200K to Millions of Tokens

In October 2023, Kimi debuted with support for a 200,000 Chinese character context window — at the time, one of the longest context windows among consumer-facing AI products. For comparison, GPT-4 offered 8K/32K tokens, and Claude 2 had 100K tokens.

By March 2024, Kimi extended its context window further to 2,000,000 Chinese characters, enabling users to upload entire books, lengthy research papers, and complete codebases for analysis and Q&A.

The name “Kimi” actually comes from Yang Zhilin’s English nickname.

Why Does Long Context Matter?

A long context window isn’t just a number to brag about — it fundamentally changes how you use the model:

• Whole-book analysis: No need to chunk or summarize — just feed the entire text and ask questions
• Long document comparison: Cross-referencing contracts, legal documents, and technical documentation
• Codebase comprehension: Understanding large codebases without RAG
• Research synthesis: Feeding a dozen papers at once for comprehensive analysis

This doesn’t conflict with RAG (Retrieval-Augmented Generation) — they’re complementary. Long context reduces dependence on external retrieval, while RAG handles knowledge at scales that exceed the context window.

Model Evolution: From Chatbot to Frontier Model

Kimi Chat (2023-2024)

The original Kimi Chat was a consumer-facing chat product focused on long-text comprehension and Chinese conversation. It quickly gained traction in the Chinese market, particularly among students, researchers, and knowledge workers.

Kimi K1.5 (January 2025)

K1.5 marked the pivotal shift from “long-text tool” to “reasoning model”:

• RL-driven reasoning: Using large-scale reinforcement learning to boost reasoning capabilities, similar to the OpenAI o1 approach
• Long chain-of-thought (Long CoT): The model can engage in deeper, longer reasoning processes
• Multimodal reasoning: Not just text — it can reason over images as well
• Math and code: Competitive with OpenAI o1 on benchmarks like AIME, MATH-500, and LiveCodeBench

K1.5’s technical report emphasized a compelling argument: you don’t need complex frameworks like Monte Carlo Tree Search — pure RL scaling alone can achieve top-tier reasoning performance.

Kimi-VL (April 2025)

A 16B MoE (3B active parameters) open-source vision-language model focused on image understanding tasks.

Kimi-Dev (June 2025)

A 72B code-specialized model (based on Qwen2.5-72B) that achieved open-source model SOTA on SWE-bench Verified.

Kimi K2 (July 2025)

K2 was Moonshot AI’s landmark open-source model, establishing the architectural direction:

• 1T total parameters, 32B active — MoE architecture with 384 experts, activating 8 per inference
• MuonClip optimizer: Combining the Muon algorithm with QK-Clip stabilization, achieving zero loss spikes across 15.5T tokens of pre-training
• 128K context window
• Modified MIT license open-source

K2 Follow-up Versions

• K2-Instruct-0905 (September 2025): Improved code capabilities, context extended to 256K
• Kimi Linear (October 2025): 48B MoE (3B active), using “Kimi Delta Attention” for more efficient inference
• K2 Thinking (November 2025): 256K context, supporting 200-300 consecutive tool calls, training cost of only ~$4.6M

Kimi K2.5 (January 2026) 🔥

K2.5 is the current flagship of the Kimi series and the biggest leap yet:

• Native multimodal: Trained from scratch on ~15 trillion mixed vision and text tokens, using MoonViT (a 400M-parameter vision encoder) — not a bolted-on vision module
• Agent Swarm: Can autonomously dispatch up to 100 sub-agents working in parallel, executing up to 1,500 tool calls, 4.5x more efficient than a single agent
• 1T MoE / 32B active, 256K context
• Modified MIT license open-source

Technical Deep Dive

MoE Architecture

Architectural details of K2/K2.5’s MoE:

Total parameters:     1.04T
Active parameters:    32B
Number of experts:    384 (8 activated per inference)
Sparsity ratio:       48:1
Attention heads:      64 (Multi-head Latent Attention)
Hidden dimension:     7168
MoE hidden dimension: 2048
Layers:               61 (including 1 dense layer)

Compared to DeepSeek-V3, K2 has more experts (384 vs 256) but fewer attention heads (64 vs 128), making different trade-offs across dimensions.

MuonClip Optimizer

This is one of Kimi’s key innovations in training stability. In traditional large model training, loss spikes (sudden surges in training loss) are a common and thorny problem. MuonClip combines:

• Muon algorithm: A token-efficient optimization method
• QK-Clip: Stabilizes gradients in the attention mechanism

Result: Zero loss spikes across the full 15.5T-token pre-training run.

The RL Path for Reasoning

Starting from K1.5, Kimi’s reasoning improvement strategy:

1. Supervised learning to train the base model
2. Large-scale agentic data synthesis pipelines to generate training data
3. Joint RL stage: combining RLVR (Reinforcement Learning with Verifiable Rewards) and Self-Critique Rubric Reward mechanisms
4. Continuous quality improvement through scaling RL compute

K2.5’s Native Multimodal

K2.5 isn’t a model with vision “bolted on” — it underwent continued pre-training on ~15T mixed vision and text tokens. MoonViT (400M parameters) is a purpose-built vision encoder that enables the model to truly “understand” images rather than merely describe them.

Kimi vs Current Mainstream Models (Early 2026)

Where Kimi Wins

• Agentic tasks: HLE with tools (44.9 vs GPT-5’s 41.7), BrowseComp (60.2 vs GPT-5’s 54.9 vs Claude’s 24.1)
• Cost-effectiveness: API pricing 4-17x cheaper than GPT-5.4, 5-6x cheaper than Claude Sonnet 4.6
• Math: AIME 2025 (96.1%), HMMT 2025 (95.4%)
• Open source: 1T-parameter model released under Modified MIT license

Where Other Models Win

• GPT-5.x: Pure reasoning, abstract problems, English writing, general knowledge (MMLU-Pro 87.1 vs K2’s 84.6)
• Claude Opus: Software engineering (SWE-bench ~80.9%), safety, writing quality (79.8 vs K2’s 73.8)
• Gemini 3 Pro: Document-intensive tasks, Google ecosystem integration, ultra-large context

API and Pricing

Moonshot AI provides the Moonshot API (platform.moonshot.ai), which is compatible with the OpenAI SDK — you can use it by simply swapping the base URL.

K2.5 Pricing

Consumer subscription: ~$8-19/month (compared to ChatGPT Plus $20/month, Claude Pro $20/month).

The model is also available through OpenRouter, Together AI, NVIDIA NIM, Hugging Face, and other platforms. The open-source version supports local deployment, with weights stored in block-FP8 format. K2 Thinking supports native INT4 quantization.

Controversy: The Distillation Incident

In February 2026, Anthropic publicly accused Moonshot AI (along with DeepSeek and MiniMax) of using approximately 24,000 fake accounts to conduct over 16 million conversations with Claude for model distillation. Moonshot AI was alleged to have conducted around 3.4 million conversations, primarily targeting agentic reasoning, tool use, code, and computer vision capabilities.

Anthropic stated that request metadata was traced back to public profiles of senior Moonshot AI employees. The incident sparked widespread debate in the industry — some viewed it as a serious intellectual property issue, while others saw it as reflecting the gray areas in AI competition.

Key Takeaways

Strengths

• Cost killer: API pricing is 1/4 to 1/17 of GPT and Claude
• Open-source strategy: 1T-parameter model released under MIT license — rare among top-tier Chinese models
• Agent Swarm: 100 parallel sub-agents is one of the most aggressive agentic architectures to date
• Native Chinese: Chinese is at the core of everything from training data to product design
• Growth velocity: K2.5’s revenue in its first 20 days exceeded all of 2025; overseas revenue has surpassed domestic; global paid users growing 170% month-over-month

Challenges

• English writing quality and software engineering benchmarks like SWE-bench still lag behind Claude and GPT
• The long-term brand reputation impact of the distillation controversy remains uncertain
• The Chinese AI market is extremely competitive — DeepSeek, Zhipu AI, and Baichuan are all fighting for the same space
• International compliance and data sovereignty concerns

Conclusion

Kimi is not “yet another Chinese ChatGPT.” Starting from long-text processing, it has undergone a complete evolution through reasoning, multimodal, and agent capabilities, becoming one of the few open-source models in the global AI market that can compete across multiple dimensions within just three years.

For developers, K2.5’s OpenAI-compatible API, highly competitive pricing, and open-source deployment options make it a seriously worth-considering choice — especially for agentic workflows and Chinese-language scenarios.

If you haven’t tried Kimi yet, I’d recommend starting with an agent task that requires extensive tool calls, or feeding it an extremely long document. These two directions are where Kimi is most compelling right now.

References

• Kimi K1.5 Technical Report
• Kimi K2 Technical Report
• Kimi K2.5 Technical Blog
• Moonshot AI Official Platform
• Kimi K2 GitHub
• Moonshot API Platform