Kimi is the conversational AI assistant and the family of large language models developed by Moonshot AI, an artificial intelligence company founded in Beijing in 2023. Among AI models & assistants, it is best known for two things: an early focus on very long context windows for reading entire books, reports and codebases, and the open-weight Kimi K2 models, which brought a trillion-parameter mixture-of-experts design to the public.
The name covers both the consumer assistant, available on the web and in mobile apps, and the underlying models offered through an API and, for the K2 generation, as downloadable weights. This dual role makes Kimi comparable to Qwen, DeepSeek and GLM, the other Chinese model families that combine a product with openly released models.
Development of the Kimi model family
Kimi has evolved from a long-document reading assistant into an agent-oriented model family. The first public assistant, launched in late 2023, distinguished itself by accepting about 200,000 Chinese characters in a single conversation, at a time when most assistants handled far less. Moonshot AI later announced support for inputs of about two million characters in limited access.
In early 2025 the company released Kimi k1.5, a reasoning model trained with large-scale reinforcement learning, and in mid-2025 it published Kimi K2 with open weights. Later releases extended K2 with longer context and a dedicated reasoning variant. The milestones below summarize the main generations.
| Release | Approximate date | Key characteristic |
|---|---|---|
| Kimi assistant | Late 2023 | Long-context document reading |
| Kimi k1.5 | Early 2025 | Reasoning trained with reinforcement learning |
| Kimi-VL | 2025 | Compact open vision-language model |
| Kimi K2 | Mid-2025 | Open-weight trillion-parameter mixture of experts |
| Kimi K2 Thinking | Late 2025 | Long-horizon reasoning with tool use |
Release names and capabilities change quickly, so exact version details should be checked against current model documentation before any procurement or benchmark comparison.
How Kimi K2 works
Kimi K2 is a sparse mixture-of-experts (MoE) transformer with about one trillion total parameters, of which only about 32 billion are active for each token. A routing network selects a small subset of expert sub-networks for every token, so the model holds a very large store of knowledge while keeping inference cost close to that of a 32-billion-parameter dense model.
Moonshot AI reported training K2 on roughly 15.5 trillion tokens with an optimizer called MuonClip, a variant of the Muon optimizer that adds a clipping step to keep attention logits stable. The company presented this as a way to avoid the loss spikes that often disrupt training of very large models. Post-training emphasized agentic data: synthetic and real tasks in which the model calls tools, reads results and plans multiple steps.
| Property | Value | What it affects |
|---|---|---|
| Total parameters | About 1 trillion | Knowledge capacity and storage size |
| Active parameters per token | About 32 billion | Inference speed and cost |
| Architecture | Mixture of experts | Efficiency through sparse routing |
| Context window | 128K, later 256K tokens | Length of input handled at once |
| Released variants | Base and instruct | Fine-tuning versus direct use |
The K2 Thinking variant adds extended internal reasoning and is designed to chain hundreds of sequential tool calls in one task, such as searching, reading and coding in a loop. It was released with quantization-aware training at INT4 precision, which reduces memory needs for self-hosting.
Kimi compared with related models
Kimi sits in the same open-weight frontier segment as DeepSeek, Qwen and GLM, but differs in emphasis. The comparison below reflects the general positioning of each family rather than any single benchmark, because benchmark rankings shift with every release.
| Model family | Developer | Distinctive emphasis |
|---|---|---|
| Kimi | Moonshot AI | Long context and agentic tool use |
| DeepSeek | DeepSeek | Training efficiency and open reasoning models |
| Qwen | Alibaba Cloud | Broad range of sizes and modalities |
| GLM | Zhipu AI | Bilingual models and agent features |
| Claude | Anthropic | Closed models focused on coding and safety |
Compared with closed frontier models such as Claude or GPT-5, Kimi K2 trades some polish and ecosystem breadth for self-hosting freedom and lower token prices. Compared with smaller open models such as Gemma or Phi, it offers much stronger general capability but requires multi-GPU server hardware to run locally.
Practical use and deployment options
Kimi can be used in three ways: through the consumer assistant, through the Moonshot AI API, or by self-hosting the open weights. The API follows a widely used chat-completions format, so many existing applications can switch to it by changing the endpoint and model name. Third-party inference providers also host K2.
- Document analysis: summarizing long contracts, financial filings or research papers in one pass.
- Agentic coding: driving coding assistants that edit files, run tests and iterate.
- Research agents: multi-step web search and synthesis with tool calls.
- Custom fine-tuning: adapting the base model to domain data under its license.
The K2 weights were published under a modified MIT license that adds an attribution requirement for very large commercial deployments. Organizations should have legal teams review the license terms before building products on the weights.
Limitations and considerations
Kimi shares the general weaknesses of large language models, including hallucinated facts and sensitivity to prompt wording, and adds several specific considerations. Self-hosting the full model demands hundreds of gigabytes of accelerator memory even with quantization.
The hosted assistant and API are operated by a company based in China, so data residency and regulatory compliance need assessment for regulated industries. The consumer assistant also follows Chinese content regulations, which can lead to refusals on politically sensitive topics. Self-hosted weights avoid data transfer but may still carry behaviors learned in training, so independent evaluation on the intended task is advisable before production use.
A plain-language summary
Kimi is a chatbot and a set of AI models from the Chinese company Moonshot AI. It became known for reading very long texts, such as a whole book, in one go.
Its newer models, called Kimi K2, are free to download and run on your own servers. They are huge, but they only use a small part of their “brain” for each word, a bit like a large company that sends each question only to the few specialists who know the answer. That makes them powerful yet cheaper to run than their size suggests.
An example from practice
A mid-sized law firm needs to compare about 40 supplier contracts, each 60 to 120 pages long, against a new internal policy. Client confidentiality rules out sending the documents to an external cloud service, so the IT team self-hosts a quantized Kimi K2 model on a rented server with eight data-center GPUs inside the firm’s own network.
An internal tool feeds each contract and the policy into a single prompt within the 128K-token context window and asks the model to list clauses that conflict with the policy. Associates then verify each flagged clause. The review, previously estimated at about three weeks of associate time, is completed in four working days, with lawyers spending their time on judgment rather than on locating clauses.