Chinese II model Kimi K1.5 mastered multimodal reasoning and surpassed Openai O1

If 2024 has become the year of ChatGPT clones, then 2025 promises to become an era of reasoning AI models, and the Chinese laboratories capture leadership in this area. Last week, a lot of noise made Deepseek with its reasoning model R1. And the other day, Moonshot AI introduced the multimodal Kimi K1.5, which overtakes in the Openai O1 tests, and costs many times less. These models are a change in the idea of ​​the “mental process” of AI.

Image source: kimi.ai

New models have gone far from the banal retelling of Wikipedia. They can do difficult problems – from solving puzzles to explanation of quantum physics. And the Kimi K1.5 has already managed to earn the title of “The First Real competitor O1.” According to experts, Kimi K1.5 is not just another AI model – this is a jump in the multimodal reasoning and reinforcement training. KIMI K1.5 from Moonshot AI combines the text, code and visual data for solving complex problems, sometimes many times superior to such industry leaders as the GPT-4O and Claude Sonnet 3.5 in key tests.

The Kimi K1.5 context window for 128 thousand tokens allows the “in one approach” model to process the amount of information equivalent to a solid novel. In mathematical tasks, the model can plan, reflect and adjust their steps for hundreds of tokens, imitating a solution to a person’s problem. Instead of re -generating complete answers, Kimi uses fragments of previous trajectories, increasing the effectiveness and reducing training costs.

Image source: Medium.com

The traditional approach, based on the principles of training with reinforcement, involves the use of complex tools, such as the search for the wood of Monte Carlo or the network of values. The Moonshot AI team abandoned them and created a simplified framework based on reinforcement learning, using the fine for the length and balance between research and operation. As a result, the developers managed to create a model that studies faster and avoids “excessive thinking” – a common mistake when AI spends computational resources on unnecessary steps.

Kimi K1.5 managed to show itself as a powerful visualization tool and simultaneous work with the text. The model can analyze diagrams, solve geometric problems and debug the code – in the Mathvista test, the model showed an accuracy of 74.9 %, combining text tips with graphic diagrams.

Researchers of Moonshot AI, instead of relying on powerful, but slow long-chain reasoning (Long-Cot), used the Long2Short method (“long-in-short”), achieving more concise and quick answers. The following methods were used for this:

• Combining models by mixing weights of long and short versions of COT.
• Sample the shortest deviation is the selection of the shortest and most correct answer from eight generated options.
• DPO optimization – teaching a model to prefer brief answers without loss of meaning.

Even with a direct comparison, the Kimi K1.5 leaves the GPT-4O and Claude Sonnet 3.5 far behind. The developers of Moonshot AI managed to optimize the process of reinforcement with:

• Hybrid deployment – joint use of GPU resources for training and withdrawal.
• Partial deployment – dividing long trajectories into controlled fragments for more effective training.
• Code sandboxes – safe media for testing the output of code, which guarantees their reliability.

According to experts, Kimi K1.5 is not just a technological breakthrough, but a look into the future of AI. Combining training with reinforcements with multimodal reasoning, this model solves problems faster, smarter and more effective.