Kimi k1.5

The large multi-modal reasoning model released by Dark Side of the Moon supports 128K ultra-long context and has leading Chinese understanding capabilities.

large language model

https://platform.kimi.com/

Visit Website

In-depth Report

Kimi k1.5 is a large multi-modal reasoning model released by Moonshot AI in January 2025, with 128K ultra-long context capabilities. This model is divided into two modes: long-chain thinking (Long-CoT) and short-chain thinking (Short-CoT). It performs well in multiple benchmark tests such as mathematical reasoning, code generation, and Chinese understanding, and some of its results surpass OpenAI o1 and GPT-4o. As a representative work of domestic large models, Kimi k1.5 has occupied a place in the global market with its high cost performance and has become a "productivity tool" for developers and users.
Moonshot AI was founded in 2023 by Yang Zhilin, a post-90s entrepreneur. The company is committed to the research and development of general artificial intelligence technology, and its core product Kimi intelligent assistant is famous for its ultra-long context processing capabilities. In January 2025, the company released the Kimi k1.5 inference model, which is divided into Long-CoT and Short-CoT versions, demonstrating powerful inference capabilities. In terms of the company's development history, Dark Side of the Moon has received great attention from the capital market since its establishment. According to LatePost, the company is about to complete a new round of financing of US$2 billion, with a post-investment valuation exceeding US$20 billion. Existing shareholders include Alibaba, Tencent and 5Y Capital.
Kimi k1.5 has the following core functions: Efficient reasoning capability Through reinforcement learning and course sampling mechanisms, the model gradually transitions from simple tasks to complex tasks, ensuring that the training process is efficient and avoids over-computation. Short-CoT mode uses a length penalty mechanism to avoid lengthy reasoning and give accurate answers in the shortest time. Ultra-long context processing supports 128K tokens ultra-long context window, which can handle long document analysis, contract review, novel creation and other tasks that require large context. Multi-modal processing supports joint training of text and visual data, has strong cross-modal reasoning capabilities, and can handle complex tasks such as chart interpretation and visual math problems. Chinese optimization has achieved leading results in Chinese benchmark tests such as CLUEWSC and C-Eval, and its Chinese understanding ability is significantly better than competing products.
The Kimi k1.5 model is currently open to API calls for free, and users can obtain the interface through the Dark Side of the Moon open platform. Enterprise users can enjoy discounts on batch calls. Please contact the official for specific pricing. Target users include individual developers and technology enthusiasts, programmers who need AI programming assistance, corporate users who need long text processing capabilities, educational institutions and student groups. Dark Side of the Moon's revenue mainly comes from Kimi paid subscriptions and API calls. According to the company's disclosure, after the Kimi K2.5 model update, the company's annual recurring revenue exceeded US$100 million in early March 2026, and further increased to over US$200 million in April.
Positive comments include strong ability to process ultra-long contexts, suitable for long document analysis; excellent Chinese understanding capabilities, especially suitable for Chinese users; excellent programming assistance capabilities, and high quality code generation; short-CoT mode has fast response speed and is suitable for daily Q&A. In terms of negative feedback, some complex mathematical reasoning tasks still have limitations; the long-chain mode has a long response time; compared with open source models such as DeepSeek-R1, the closed-source mode is not flexible enough. Usage scenarios include code writing and debugging assistance, long document analysis and summary, mathematical problem solving, Chinese and English translation, creative writing and copywriting generation.
In terms of media opinions, according to LatePost, Kimi has become one of the benchmark products for domestic large model startups. Geek Park comments pointed out that Kimi has demonstrated a breakthrough in reasoning capabilities of domestic large models. Zhihu users discussed that the technical route of Kimi k1.5 is innovative and the reinforcement learning training method deserves attention. Expert analysis believes that the success of Kimi k1.5 lies in the use of a simple and efficient RL framework, the Long2Short method to realize knowledge transfer, and the profound accumulation in the field of Chinese science and technology. In terms of competitive product landscape, main competitors include OpenAI o1, DeepSeek-R1, QwQ-32B, etc.
In terms of technical controversy, some people believe that although the benchmark test results of Kimi k1.5 are good, the performance in actual applications still needs more verification. Market risks include fierce competition among large models and rapid technology iterations, the rise of open source models putting pressure on closed source business models, and intensifying competition in the Chinese market. Potential problems include computational cost pressure brought by long context, technical challenges of multi-modal fusion, and geopolitical risks of global market expansion.
It is suitable for programmers and analysts who need to process long documents, Chinese content creators and educators, development teams who need AI programming assistance, and users who have high requirements for Chinese understanding ability. Who it is not suitable for includes real-time application scenarios that require extremely high response speed, enterprise users who require completely open source solutions, and individual users who are extremely price-sensitive. Alternatives include the OpenAI o1 international mainstream inference model, the DeepSeek-R1 open source inference model that can be deployed by yourself, and the QwQ-32B open source long chain inference model.
Kimi k1.5 is a high-performance reasoning model launched by Dark Side of the Moon, which has demonstrated leading capabilities in mathematical reasoning, Chinese understanding and other fields. With its 128K ultra-long context and Short-CoT efficient reasoning capabilities, this model has become one of the benchmark products for domestic large models. As the company completes a new round of financing, the commercialization process of Kimi's products will be further accelerated and it is expected to occupy a more important position in the global market in the future. For Chinese users, Kimi k1.5 is a reasoning tool worth trying, especially in application scenarios that require long context processing and Chinese understanding. It is recommended that individual developers use the API free quota to experience it, and enterprise users can contact the official to obtain customized solutions.

User Reviews

BCook_77

—

The Short-CoT mode of Kimi k1.5 is really strong, not to mention its fast response speed, and its mathematical reasoning ability is actually better than GPT-4o!

CThompson759

—

128K context is really great. It is very convenient to use to process long papers. You can summarize it without dividing it into sections.

JudyRichardson_Plus

—

The models trained by reinforcement learning are indeed different. The reasoning process is clearer and the thinking is more like humans.

RonaldRussell520583

—

The response is much faster than DeepSeek-R1. Although the performance is similar, the experience is much better.

PFisher_66

—

Chinese understanding ability yyds! I tried several classical Chinese comprehension questions and answered them accurately, showing a good understanding of the cultural background.

LoganButler_66

—

The Long2Short method has something special. It transfers the ability of long-chain thinking to the short-chain model, taking into account both efficiency and effect.

heavytiger762

—

The free API quota is enough, so it’s no problem to start a small project, but I don’t know if there will be any charges in the future.

علی رضاسلطانی نژاد

—

The programming assistance ability is not as good as Claude, but it is very easy to use in Chinese scenes. Each has its own advantages.

Natalie_Vasquez16884

—

The long chain mode sometimes takes too long to think about, but the short chain mode is just right and is completely sufficient for daily Q&A.

KathrynCooper

—

The company has raised 20 billion yuan in financing, and its products are still free. I hope it can remain so.

Doris.Robinson_77

—

There is something about multi-modal capabilities. Last time I sent a screenshot and asked him to help me analyze the code, and he directly gave a detailed explanation.

秋叶_8

—

MathVista test scores are higher than o1, and visual reasoning is really strong.

Amy_Morales_99

—

It is much easier to use than Tongyi Qianwen, especially in long text processing scenarios.

DouglasLopez_77

—

The AIME 2024 mathematics test has a pass rate of 77.5%, which is higher than OpenAI o1. The domestic model is outstanding!

Tyler_Bell_2022

—

After using it for a while, I feel that it is very easy to use, but I hope that more parameter scale options can be opened.

Natalie_Baker

—

You will still be stumped by the 24-point question in the actual test, and there is room for improvement in your reasoning ability.

BrendaLee007

—

The CLUEWSC Chinese comprehension test scored 91.7, which is nearly 4 points higher than o1. Chinese users are ecstatic.

翡翠93

—

The code generation capability is not as good as the specialized Coding model, but the comprehensive capability is strong.

TokenMasterWerner

—

The C-Eval test score is 88.3 points, which is a big lead, and the Chinese test ability is unparalleled.

Victoria.Green_2024

—

The Dark Side of the Moon is winning this time, and has done well in both technological breakthroughs and commercialization.

Kimi k1.5

In-depth Report

User Reviews

Similar Products