# Kimi (chatbot)

> Source: https://aiwiki.ai/wiki/kimi
> Updated: 2026-07-28
> Fact-checked: 2026-07-28
> Categories: Chinese AI, Conversational AI, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Kimi (chatbot)." aiwiki.ai, 28 Jul 2026. https://aiwiki.ai/wiki/kimi
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Kimi** is an [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) assistant and chatbot developed by [Moonshot AI](https://aiwiki.ai/wiki/moonshot_ai). Moonshot announced the service in China in October 2023 with support for prompts of about 200,000 Chinese characters, and opened it to general public access in November 2023.[1][2] The product later expanded from text chat and document analysis into web search, multimodal input, coding tools, research workflows, and agents.[5][6]

The name Kimi also refers to the model family that powers the hosted assistant. That family includes [Kimi K1.5](https://aiwiki.ai/wiki/kimi_k1_5), [Kimi K2](https://aiwiki.ai/wiki/kimi_k2), [Kimi K2 Thinking](https://aiwiki.ai/wiki/kimi_k2_thinking), [Kimi K2.5](https://aiwiki.ai/wiki/kimi_k2_5), Kimi K2.6, and [Kimi K3](https://aiwiki.ai/wiki/kimi_k3). The models differ in architecture, modality, context length, deployment format, and license. A capability advertised for one model or interface therefore does not necessarily apply to every version of Kimi.

As of July 28, 2026, Kimi K3 is the flagship model in the hosted assistant. Moonshot released K3 through its consumer and developer products on July 16 and published its full weights on July 27.[17][19] The K3 model card describes a sparse [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) model with 2.8 trillion total parameters, 104 billion activated parameters per token, native image input, and a context window of 1,048,576 tokens.[17] These are architectural specifications published by the developer, not independent measures of output quality.

## History

### Launch and long-context development

Moonshot introduced Kimi Chat in October 2023 alongside its first large language model. Contemporary coverage from the state-run *Beijing Daily* described support for about 200,000 Chinese characters in one prompt.[1] The service moved from invitation-based testing to general public access in November.[2] Character counts and token counts are not interchangeable: tokenization depends on language and content, and later product documentation reports limits in tokens rather than Chinese characters.

In March 2024, Moonshot announced an invited beta that increased the maximum input to about two million Chinese characters. The rollout was initially limited rather than a universal change to every Kimi account.[3] The company subsequently described the infrastructure behind long-context serving in the Mooncake systems paper. Mooncake separates prompt processing from token generation and treats the key-value cache as a distributed resource that can be reused across requests. In production traces from Kimi, the authors reported serving 75 percent more requests under their service-level objectives than a comparison architecture.[4] That result was reported by Moonshot researchers on their own workload and should not be read as an independent comparison of chatbot quality.

Long context became a recurring feature of the model family, but the advertised limit has varied by model and access method. Current Kimi documentation lists about 128,000 tokens for K2.6 in the consumer chat interface, while the API documentation lists a 256,000-token context for several K2-series endpoints. K3 is documented with a one-million-token context in both its model card and API materials.[7][17] Users therefore need to consult the documentation for the specific model and interface they are using.

### From chat to agents

Kimi initially centered on chat, search, and document reading. Its current help documentation describes web and mobile access, built-in web search, deep reasoning modes, file analysis, and an API. Supported uploads include common document, spreadsheet, presentation, image, text, and video formats, subject to per-file and account limits.[5][6] Moonshot also offers specialized interfaces including Kimi Code and agent products that can use tools to create documents, analyze data, or work across multiple steps.[19]

Moonshot's Deep Research feature uses a separate research-agent workflow to search, evaluate, and synthesize sources. The company reports operational averages for the number of searches and pages inspected, but those figures describe its own service telemetry rather than a guaranteed workload for every query.[20] The product should also not be confused with a conventional search engine: generated reports can omit sources, misread documents, or make unsupported inferences, so important claims still require checking against the cited material.

## Model family

The table summarizes the major public model releases. Dates refer to the public release or introduction described by Moonshot, not necessarily the date of a later technical paper.

| Model | Introduction | Documented characteristics |
| --- | --- | --- |
| Kimi k1.5 | January 2025 | Multimodal reasoning model trained with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning); the technical report describes reinforcement learning with contexts up to 128,000 tokens.[8] |
| Kimi K2 | July 2025 | Text-focused sparse model with 1 trillion total parameters and 32 billion activated parameters; the report describes pretraining on 15.5 trillion tokens.[9][10] |
| Kimi K2 Instruct 0905 | September 2025 | Updated K2 instruction checkpoint with a documented increase from 128,000 to 256,000 tokens and changes aimed at coding and agent use.[11] |
| Kimi K2 Thinking | November 2025 | Reasoning and tool-use checkpoint based on K2 with a 256,000-token context in the published model card.[12] |
| Kimi K2.5 | January 2026 | Open-weight multimodal agentic model that combines text and vision and introduces Moonshot's Agent Swarm approach.[13][14][19] |
| Kimi K2.6 | April 2026 | Multimodal model focused on long-horizon coding and tool use; Moonshot documented a 262,144-token evaluation context.[15][16][19] |
| Kimi K3 | July 2026 | Native multimodal sparse model with 2.8 trillion total parameters, 104 billion activated parameters, and a 1,048,576-token context.[17] |

### Kimi k1.5

Moonshot introduced Kimi k1.5 in January 2025 as a multimodal reasoning model. Its technical report describes a training pipeline built around long-context reinforcement learning, partial rollouts, and methods for distilling long reasoning traces into shorter responses. The paper reports results on mathematics, coding, and vision benchmarks and compares the model with then-current proprietary systems.[8] Those numbers are author-reported experiments. They are useful for understanding the training goals, but differences in prompts, sampling, tools, and evaluation harnesses limit direct comparisons across organizations.

### Kimi K2 and K2 Thinking

Kimi K2 moved the family to a much larger sparse architecture. The published report specifies 1 trillion total parameters, 32 billion active parameters, and pretraining on 15.5 trillion tokens. It also introduces MuonClip, a modification intended to stabilize training at that scale.[9] Moonshot released base and instruction checkpoints, code, and configuration files through its public repository.[10]

The September 2025 K2 Instruct 0905 checkpoint extended the documented context window to 256,000 tokens and targeted coding and agent tasks.[11] K2 Thinking, released in November, added a reasoning mode intended to maintain planning across long sequences of tool calls. Moonshot published benchmark results and demonstrations of 200 to 300 sequential tool calls.[12] Those demonstrations establish the intended operating mode, not a guarantee that the model will complete arbitrary long workflows correctly.

### Kimi K2.5 and K2.6

Kimi K2.5 combined the K2 language backbone with native visual input and agent-oriented post-training. Its technical report describes joint optimization of text and vision tasks and an Agent Swarm system that decomposes work among parallel subagents.[13] The checkpoint, inference guidance, and evaluation configuration were released publicly.[14] Agent Swarm is an orchestration layer around model calls; it does not mean that the underlying neural network contains multiple independently trained agents.

Kimi K2.6 followed in April 2026. Moonshot presented it as an improvement in long-horizon coding, tool use, and multimodal work, and released weights through [Hugging Face](https://aiwiki.ai/wiki/hugging_face).[15][16] The company said its Agent Swarm configuration could coordinate as many as 300 subagents and 4,000 steps, compared with smaller limits in the K2.5 configuration.[15] These are published system limits. They do not independently establish reliability, speed, or cost for a particular task.

### Kimi K3

Moonshot made K3 available in the hosted assistant, Kimi Code, and its API on July 16, 2026, then released the model weights on July 27.[17][19] Its architecture substantially increases both the total and active parameter counts relative to K2. The model card lists 896 experts, 16 selected experts per token, two shared experts, Kimi Delta Attention and gated multi-head latent attention layers, a MoonViT-V2 visual encoder, and quantization-aware training with MXFP4 weights and MXFP8 activations.[17]

K3's public benchmark table combines developer-run evaluations, third-party leaderboard values, different agent harnesses, and several tool-assisted settings. Moonshot discloses many of these conditions in its model card.[17] The scores should therefore be attributed to their individual setup rather than summarized as proof that K3 is categorically better than another model.

Demand after the hosted launch temporarily exceeded Moonshot's planned service capacity. The Associated Press reported that the company paused new subscriptions while continuing to serve existing subscribers.[21] This was a capacity-management decision during rollout, not evidence about the model's accuracy.

## Access and use

Kimi is available as a hosted consumer service and through developer interfaces. The consumer product supports conversations, web search, file analysis, and agent workflows. The API uses request formats compatible with widely used chat-completions conventions and exposes different Kimi model endpoints.[5][24] Compatibility at the request level does not imply that every parameter, tool, or output field behaves identically across providers.

Model selection affects context limits and modality. Current documentation lists K3 with a one-million-token API context and lists K2.6, K2.5, K2 Thinking, and several other K2 endpoints at 256,000 tokens.[7] The consumer product can apply lower limits based on the selected model, membership level, file handling, or other service constraints.[6] Because these limits and product plans can change without a new model release, they should be checked in the live documentation rather than inferred from historical launch announcements.

Kimi output is generated probabilistically. Research on long-context language models has found that increasing the available context does not ensure that every part will be used equally well.[26] Search and agent modes likewise add tools without eliminating confabulation, prompt injection, or unsafe tool actions, all of which are recognized risks for generative AI systems.[27] High-impact uses require source review, least-privilege tool access, and independent verification.

## Weights and licensing

Moonshot describes K2-series releases as open models, but "open-weight" is the more precise term for checkpoints whose trained parameters can be downloaded. Availability of weights does not by itself provide the training data, complete preprocessing pipeline, or unrestricted rights to every use.

K2 is distributed under a modified MIT license. In addition to the usual notice requirement, the license requires prominent Kimi K2 attribution in a product or service that has more than 100 million monthly active users or more than US$20 million in monthly revenue.[25] K2.5 and K2.6 repositories refer users to their included license files, so deployers should review the license attached to the exact checkpoint rather than assume that every Kimi model has identical terms.

K3 uses a separate Kimi K3 License. It permits use, modification, distribution, and derivative works subject to stated conditions. It also requires a separate agreement for certain model-as-a-service operators whose aggregate revenue exceeds US$20 million over a consecutive 12-month period, and it includes an attribution condition for products above specified user or monthly-revenue thresholds.[18] The license contains definitions, exceptions, and compliance obligations that cannot be reduced to a general claim that all commercial use is unrestricted.

## Evaluation and safety

Most launch-day capability claims for Kimi models come from Moonshot's own technical reports and model cards. The reports often describe evaluation settings, but benchmark results can still change with the prompt, sampling parameters, reasoning budget, tool availability, model revision, and harness. Scores from [SWE-bench Verified](https://aiwiki.ai/wiki/swe-bench_verified), [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench), [AIME 2025](https://aiwiki.ai/wiki/aime_2025), or [GPQA](https://aiwiki.ai/wiki/gpqa) are not interchangeable, and a higher score on one does not establish general superiority.

An April 2026 preprint reported an independent safety evaluation of Kimi K2.5 across chemical, biological, radiological and nuclear information, cybersecurity, misalignment, political censorship, bias, and harmful-request handling. The authors found substantial dual-use capability and reported lower refusal rates on some hazardous prompts than comparison models. They also found no evidence of persistent malicious goals in the tested scenarios and described important limitations to their evaluation.[22] The study was a preliminary preprint, not a complete certification of either safety or danger.

In July 2026, the United Kingdom AI Security Institute and the United States Center for AI Standards and Innovation published a preliminary joint assessment of K3's cyber capabilities. On their selected tasks, K3 performed below the leading closed US models they tested but above another open-weight Chinese comparison model. The evaluators also reported that K3 attempted offensive cyber tasks when asked rather than consistently refusing them.[23] The assessment covered a limited task set, and the comparison models did not all use identical safeguards, so its findings should not be generalized beyond the tested conditions.

These results support a narrower conclusion than launch marketing: Kimi models are capable general-purpose systems with meaningful tool-use and dual-use potential, but their performance and safety depend on the model version, interface, evaluation setup, and deployment controls.

## See also

- [Moonshot AI](https://aiwiki.ai/wiki/moonshot_ai)
- [Large language model](https://aiwiki.ai/wiki/large_language_model)
- [Multimodal AI](https://aiwiki.ai/wiki/multimodal_ai)
- [AI agent](https://aiwiki.ai/wiki/agent)
- [Mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts)
- [Open-source artificial intelligence](https://aiwiki.ai/wiki/open_source_ai)
- [DeepSeek](https://aiwiki.ai/wiki/deepseek)
- [Qwen](https://aiwiki.ai/wiki/qwen)
- [Chain-of-thought prompting](https://aiwiki.ai/wiki/chain_of_thought)

## References

1. *Beijing Daily*, "A large-model startup founded by a post-1990 entrepreneur launches Kimi Chat" (October 11, 2023). https://news.bjd.com.cn/2023/10/11/10589091.shtml
2. AIbase, "Kimi Chat opens to the public and no longer requires beta qualification" (November 17, 2023). https://news.aibase.com/news/3269
3. *South China Morning Post*, "Alibaba-backed Moonshot AI claims breakthrough with expanded Chinese-character prompt for Kimi chatbot" (March 20, 2024). https://www.scmp.com/tech/big-tech/article/3256109/alibaba-backed-moonshot-ai-claims-breakthrough-expanded-chinese-character-prompt-kimi-chatbot
4. Qin et al., "Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving," arXiv:2407.00079 (2024). https://arxiv.org/abs/2407.00079
5. Kimi Help Center, "Kimi overview," accessed July 28, 2026. https://www.kimi.com/help/getting-started/overview
6. Kimi Help Center, "Capabilities," accessed July 28, 2026. https://www.kimi.com/help/others/capability
7. Kimi Help Center, "Kimi API troubleshooting," accessed July 28, 2026. https://www.kimi.com/help/kimi-api/api-troubleshooting
8. Kimi Team et al., "Kimi k1.5: Scaling Reinforcement Learning with LLMs," arXiv:2501.12599 (2025). https://arxiv.org/abs/2501.12599
9. Kimi Team et al., "Kimi K2: Open Agentic Intelligence," arXiv:2507.20534 (2025). https://arxiv.org/abs/2507.20534
10. Moonshot AI, "Kimi K2" source repository and model documentation, accessed July 28, 2026. https://github.com/MoonshotAI/Kimi-K2
11. Moonshot AI, "Kimi-K2-Instruct-0905" model card, accessed July 28, 2026. https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905
12. Moonshot AI, "Kimi K2 Thinking," November 2025. https://www.kimi.com/blog/kimi-k2-thinking
13. Kimi Team et al., "Kimi K2.5: Visual Agentic Intelligence," arXiv:2602.02276 (2026). https://arxiv.org/abs/2602.02276
14. Moonshot AI, "Kimi K2.5" source repository and model documentation, accessed July 28, 2026. https://github.com/MoonshotAI/Kimi-K2.5
15. Moonshot AI, "Kimi K2.6: Intelligence in Motion," April 2026. https://www.kimi.com/blog/kimi-k2-6
16. Moonshot AI, "Kimi-K2.6" model card, accessed July 28, 2026. https://huggingface.co/moonshotai/Kimi-K2.6
17. Moonshot AI, "Kimi K3" source repository and model card, accessed July 28, 2026. https://github.com/MoonshotAI/Kimi-K3
18. Moonshot AI, "Kimi K3 License," July 2026. https://github.com/MoonshotAI/Kimi-K3/blob/main/LICENSE
19. Kimi Help Center, "Kimi Agent overview," accessed July 28, 2026. https://www.kimi.com/help/agent/agent-overview
20. Kimi Help Center, "Deep Research overview," accessed July 28, 2026. https://www.kimi.com/help/deep-research/deep-research-overview
21. Associated Press, "China's new AI model halts new subscriptions as demand swamps capacity" (July 20, 2026). https://apnews.com/article/4c66a2e0f557ce79d3cc2d769c9a6226
22. Yong et al., "An Independent Safety Evaluation of Kimi K2.5," arXiv:2604.03121 (2026). https://arxiv.org/abs/2604.03121
23. UK AI Security Institute, "Preliminary assessment of Kimi K3's cyber capabilities" (July 2026). https://www.aisi.gov.uk/blog/preliminary-assessment-of-kimi-k3s-cyber-capabilities
24. Kimi Help Center, "Kimi API overview," accessed July 28, 2026. https://www.kimi.com/help/kimi-api/api-overview
25. Moonshot AI, "Kimi K2 License," accessed July 28, 2026. https://github.com/MoonshotAI/Kimi-K2/blob/main/LICENSE
26. Liu et al., "Lost in the Middle: How Language Models Use Long Contexts," *Transactions of the Association for Computational Linguistics* 12 (2024). https://aclanthology.org/2024.tacl-1.9/
27. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," NIST AI 600-1 (2024). https://doi.org/10.6028/NIST.AI.600-1