AI without logic, but wisely: what is interesting Kimi-K2, which surpassed GPT-4

The Chinese company Moonshot AI introduced the Kimi-K2-a new large language model with open weights that can compete with commercial analogues, such as GPT-4.1 and Claude Sonnet 4. The model does not have a specialized module of logical output, but demonstrates high results in programming, mathematics and multi-language tasks.

Nikita Sushko, a junior researcher at the Asseting NLP group of the Airi Institute-spoke about the key features of the Kimi-K2 and its potential. Kimi-K2 is built on the architecture of MIXTURA of Experts (MOE), where instead of one deep model, many simpler “experts” are used. This allows you to accelerate the generation of answers and increase the amount of knowledge that the model can apply.

The key innovation of the Kimi-K2 is not architecture, but its size: the previous largest open model has 671 billion parameters (Deepseek V3), while the Kimi-K2 already has 1 trillion parameters.

Despite a huge number of parameters, the speed of the model remains comparable to its predecessors due to the optimized structure.

On tests related to the code, Kimi-K2 shows high results. For example, on Swe-Bench Verified, where models need to be found and corrected in real projects, it scored 65.8%, bypassing GPT-4.1 (54.6%) and losing only the Claude Sonnet 4.

According to statistics, 29% of the requests for ChatGPT are somehow related to the code, so this is one of the key uses of existing language models.

The model also leads to LiveCodebench (53.7%) and OJBENCH (27.1%), demonstrating the ability to solve interactive and competitive programming problems.

Moonshot AI positions Kimi-K2 as a model for agent applications-it can execute commands, call external tools, generate and correct the code, and solve multi-step problems.

When learning Kimi-K2, a lot of effort was invested in the creation of dataset with instruments.

For example, the model can analyze salaries, conduct a statistical assessment and create an interactive HTML page with recommendations-all in one process.

Moonshot AI, founded in 2023, received serious support from Chinese giants Tencent, Alibaba and Gaorong Capital. The appearance of Kimi-K2 strengthens the position of China in the AI race, where there are already such models as Deepseek and QWEN.

Users competition only on hand. Kimi-K2 is quite cheap compared to many competitors.

The openness of the scales allows you to unfold the model at its own capacities, which makes it attractive to developers and companies.

The creators of the Kimi-K2 plan to add a module of the logical conclusion, which can further increase its effectiveness. In the meantime, the model has already proved its competitiveness in key tasks – from programming to complex mathematical calculations.