Moonshot AI has moved Kimi k1.5 from a technical announcement into a product more people can try. The company’s latest reasoning model is now available through Kimi.ai, with access to its full feature set offered for free and without usage limits.
The launch matters because Kimi k1.5 is being positioned in the same competitive field as OpenAI’s o1 and DeepSeek-R1. It is also a multimodal model, meaning it can work with both text and images rather than text alone.
What Kimi k1.5 Offers on the Web
The free web version brings Kimi k1.5 to a broader audience through Kimi.ai. The model now works in English too, although Moonshot AI says that language support is still being fine-tuned.
According to Moonshot AI’s announcement, users can access the full feature set without usage limits. The system can search the web in real time across more than 100 websites, process up to 50 files at once, and use improved reasoning and image understanding capabilities.
There is still a sign-up requirement. To get started, users need either a Chinese or US phone number. Google sign-in is described as coming soon.
Those details make Kimi k1.5 more than a benchmark story. A reasoning model that can take documents, search current websites, and interpret images is aimed at tasks where users need more than a short chatbot answer. It can be used as a single interface for comparing sources, reading files, and working through questions that involve visual or textual evidence.
Why the Model Is Being Compared With o1 and DeepSeek-R1
Kimi k1.5 arrived after DeepSeek-R1, adding another Chinese reasoning model to a fast-moving category. Moonshot AI describes the model as strong on complex reasoning tasks, and the company’s technical report presents results that match or exceed leading models including OpenAI’s o1 and DeepSeek-R1.
The model comes in two versions. One is designed for detailed reasoning, known as long-CoT. The other is built for concise answers, known as short-CoT.
The difference is practical. The long-CoT version works through problems step by step, while the short-CoT version is meant to respond more briefly. In several benchmarks, the model performs as well as or better than GPT-4o and Claude 3.5 Sonnet, according to the source article’s summary of the company’s technical report.
The multimodal capability also separates Kimi k1.5 from DeepSeek-R1. Kimi k1.5 can process text and images, which lets it reason across different kinds of input. The model scores particularly well on multimodal benchmarks such as MathVista and MMMU.
Benchmarks are still only one part of the picture. The important test is whether the same strengths hold up in real use, where prompts, files, images and user expectations are less controlled than benchmark tasks.
How Moonshot AI Trained Kimi k1.5
The development process began with standard pre-training on a massive dataset of text and images. That stage was used to build the model’s general language and visual understanding.
After pre-training, Moonshot AI fine-tuned the model on a smaller, carefully selected dataset using SFT. For problems with clear right or wrong answers, such as math questions, the team used rejection sampling. That means multiple answers were generated, and only correct answers were kept for training.
The team also created extra training data that showed detailed step-by-step reasoning. That matters because reasoning models are often judged not only by the final response, but by whether they can work through a complex task in a structured way.
The final phase used reinforcement learning. Moonshot AI’s approach focused on the final outcome rather than using value functions to judge intermediate steps. The company argues that this gives the model more room to explore different routes to a correct answer.
To keep answers from becoming inefficient, the team added a penalty for overly long responses. That design choice connects directly to the split between long-CoT and short-CoT: strong reasoning is valuable, but length and compute cost still matter.
Making Reasoning Shorter and More Efficient
Detailed reasoning can improve results, but it also requires more computing power. Moonshot AI therefore worked on ways to transfer knowledge from longer reasoning models into models that answer more briefly.
The techniques included model fusion and Shortest Rejection Sampling. Shortest Rejection Sampling selects the most concise correct answer from multiple attempts, which helps preserve accuracy while reducing unnecessary length.
The team also found that increasing context length up to 128k tokens consistently improved performance. Longer context gives the model more room to handle complex reasoning, especially when a task requires more information to stay in view at once.
Moonshot AI’s research also suggests that effective reasoning models do not need complicated components such as Monte-Carlo Tree Search. That finding is described as similar to what DeepSeek-R1’s developers discovered.
The broader lesson is clear: reasoning performance is not only about making a model think longer. It is also about deciding when detailed reasoning is useful, how to compress that ability into shorter answers, and how to avoid spending compute on unnecessary output.
Moonshot AI’s Bigger Push
Kimi k1.5 is part of Moonshot AI’s larger growth story. The company was founded in 2023 and secured over $1 billion in funding led by Alibaba in February 2024, reaching a $2.5 billion valuation.
By August, that valuation had grown to $3.3 billion after additional investment from Tencent and Gaorong Capital. The source article notes that Kimi k1.5 will power the company’s ChatGPT competitor.
The web release changes the immediate picture for users. Earlier, the article said Moonshot AI had not yet made the models publicly available. With the January 25, 2025 update, Kimi k1.5 is now available through Kimi.ai, giving users a direct way to test the model’s reasoning, web search, file handling and image understanding in practice.
For the AI market, that makes Kimi k1.5 another model to watch in the reasoning race. For users, the more concrete question is simpler: whether its free web access, multimodal input, real-time web search and file processing can deliver useful answers outside controlled benchmark settings.