Chinese startup Moonshot AI’s Kimi k1.5 model outperforms OpenAI-o1 after DeepSeek-R1

With China taking center stage in international discussions, the AI arms race is intensifying. The AI community was just beginning to comprehend the potential of DeepSeek’s DeepSeek-R1, which competes with the OpenAI-o1 model when a new player emerged asserting that it was on par with or even better than o1. DeepSeek marked the official start of the era of Chinese AI models, and the Kimi K1.5 has since surpassed OpenAI’s GPT-4o and Claude 3.5 Sonnet on several important benchmarks. 

Also Read: Alibaba Reveals Improved AI Model, Says It Beats Competitor DeepSeek-V3

Kimi K1.5: What is it?

The most recent model from Beijing-based Moonshot AI, an AI startup, is the Kimi K1.5. According to reports, the recently released Kimi performs on par with or better than the OpenAI-01 model. The O1 model can handle more complicated challenges since it is built to think more deeply before acting. In contrast to DeepSeek-r1, Kimi is multimodal and has reportedly excelled o1 in domains including arithmetic, coding, and the comprehension of text and visual inputs like images and videos. Kimi is constructed at a fraction of the expense of building frontier AI models in the United States, much like DeepSeek’s AI models. With its release on Kimi.ai, the Kimi K1.5 has been heralded as the first real rival to the O1.

What makes Kimi unique?

According to reports, Kimi K1.5 is more than simply an AI model; it is being hailed as a significant advancement in multimodal reasoning and reinforcement learning (RL). The approach can handle complicated problems by combining text, code, and visual data. The model has outperformed Claude Sonnet 3.5 and GPT-4o according to the benchmarks. The Kimi team has published a comprehensive study outlining the model’s obstacles and how it has accomplished its breakthrough. 

Also Read: CRED becomes first fintech to include RBI’s digital currency into its e₹ wallet beta edition

Kimi k1.5 is a multi-modal big language model that was built by reinforcement learning techniques. Given its ability to handle many data types, Kimi is a flexible model with a wide range of uses. Kimi learns by exploration and rewards, whereas traditional AI models rely on static information. It is said that this procedure improves its capacity to analyze and resolve challenging issues. 

Kimi K1.5 outperformed OpenAI’s GPT-4 variations with a score of 96.2 on the MATH 500. It received a score of 77.5 on the math benchmark AIME and the 94th percentile on the coding-related Codeforces test. Additionally, the model outperformed Claude 3.5 Sonnet and GPT-4 by up to 550 percent on a few of the benchmarks. In terms of reasoning and problem-solving skills, Kimi outperforms its US counterparts, such as the GPT-4 and Claude models. Complex mathematics and long-context jobs are handled effectively by the model. However, since AI businesses do the tests independently and release the results, the validity of benchmark test scores is sometimes questioned.

Also Read: DeepSeek surpasses ChatGPT to lead the US Apple App Store

How does it operate?

As previously stated, Kimi uses reinforcement learning approaches to increase its decision-making skills; it advances by investigating and honing solutions. To strengthen its logic, the model employs the Chain of Thought technique, which divides difficult issues into manageable chunks. According to the study article, the model can comprehend and produce responses based on a large amount of data since it can handle extended context windows—up to 128k tokens. Kimi can be used for activities like text-image analysis and problem-solving that require visual input because it can process and reason across text and images. In terms of efficiency, the model employs techniques called length penalties and partial rollouts, which avoid lengthy responses and reuse prior outputs.