News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgwavKimi

Kimi

A Chinese chatbot and large language model family by Moonshot AI.

Moonshot AI launched Kimi in October 2023 as a chatbot with massive context handling. The series expanded into open weights models like the 2.8 trillion parameter K3, supporting text, image, and video inputs for global users via web and API.

Share on WhatsApp

  • Released in October 2023 by Moonshot AI.
  • K3 model has 2.8 trillion parameters.
  • Supports up to two million character context.
  • Uses Stable LatentMoE architecture.
  • Faced distillation accusations from Anthropic.

What it is

Kimi is a suite of large language models and an accompanying chatbot created by the Chinese firm Moonshot AI. First released in October 2023, it gained attention for handling extremely long input contexts. The latest flagship, Kimi K3, arrived in July 2026 with 2.8 trillion parameters. It processes text, images, and video while generating text, audio, and visual outputs. The system supports agentic workflows through features like Swarm and OK Computer, allowing autonomous task execution across multiple sub-agents.

How it came about

Moonshot AI founded the project in March 2023, launching the closed beta chatbot by October. Early versions focused on context length, reaching 128,000 tokens initially and later expanding to two million characters. Public release followed in November 2023. Subsequent updates introduced multimodal capabilities, coding tools, and research agents. The model lineage progressed from K1.5 in early 2025 to the massive K3 series in mid-2026, with specialized variants for finance and code appearing later that year.

Inside the model

The architecture employs a mixture of experts structure, specifically using Stable LatentMoE with hundreds of routed experts but only a small fraction active per token. This design reduces computational load while maintaining capacity. The models utilize Attention Residuals, a mechanism where attention operates over the accumulated residual stream rather than just the previous layer. Feedforward networks use SiTU-GLU modules, replacing standard linear maps with tanh functions to improve stability and performance during training and inference.

What it can do

Kimi models excel at processing vast amounts of data, supporting up to two million character contexts. They handle multimodal inputs including images and video. Agentic features allow the creation of websites and slides from simple prompts. Benchmarks show strong performance in mathematics, coding, and reasoning. K3 ranked third on the Artificial Analysis leaderboard in July 2026, though it led in specific front-end web development tests. The system can output text, audio, and images simultaneously.

How to use it

Users access Kimi via web applications, iOS and Android apps, and an API. Open weights are available for certain versions under custom licenses. These licenses impose conditions on commercial use, requiring revenue sharing for providers exceeding twenty million dollars annually. Attribution is mandatory for products with high traffic or revenue. The Kimi Code agent provides command-line access for software development tasks directly within the ecosystem.

Safety and testing

The company has implemented various safety measures and evaluations for its models. However, significant controversy arose in September 2026 when Anthropic accused Moonshot AI of secretly routing user requests to Claude models to distill knowledge. This allegation suggested covert data usage beyond the stated terms. Moonshot faced further criticism regarding the distillation claims for K3, with external officials questioning the origins of the model's capabilities relative to existing proprietary systems.

Reception and debate

Kimi received notable praise for its context window capabilities, being among the first to support such lengths publicly. Its multimodal and agentic features drew interest from developers and researchers. Commercial adoption grew rapidly, with monthly active users surpassing thirty-six million. Legal disputes emerged, particularly involving Anthropic's accusations of distillation. These incidents sparked debate about transparency in model training and the ethical use of competitor data in the open weights community.

Why it matters

Kimi influenced the open weights landscape by demonstrating the viability of massive parameter counts with efficient expert routing. Its licensing model challenged traditional open source norms by introducing revenue-sharing requirements. US-based companies utilized Kimi as a foundation for their own models, indicating its technical competence. The controversies surrounding distillation highlighted tensions between open development and proprietary protection, shaping future discussions on model provenance and fair use in AI.

Questions people ask about Kimi

What is the maximum context length supported by Kimi?

Kimi supports a context window of up to two million characters. This allows the model to process extremely long documents or conversations without losing information, a feature that distinguished it from earlier AI assistants upon its initial public release.

How does the Kimi K3 license work?

Kimi K3 uses a custom license requiring revenue sharing for inference providers earning over twenty million dollars annually. Companies with high traffic must also display attribution. This approach blends open weight availability with commercial restrictions to protect the developer's interests.

What is the Stable LatentMoE architecture?

It is a mixture of experts variant used in Kimi K3. It routes inputs to hundreds of experts but activates only sixteen per token. This sparsity reduces computation while maintaining model quality, making the large parameter count feasible for practical deployment.

Has Kimi been used by other companies?

Yes, Kimi models served as the basis or for post-training in US models like Inkling and Cursor Composer 2. This indicates that other developers view Kimi as a valuable foundation for building their own specialized AI systems.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Kimi (AI)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-16.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.