Kimi k1.5
Multimodal LLM trained with reinforcement learning for reasoning
Kimi k1.5 is a multimodal large language model developed by the Kimi Team. Released in early 2025, it uses reinforcement learning to scale training data and improve reasoning. The model achieves high scores on mathematics and coding benchmarks, matching or exceeding other leading systems.
- Uses reinforcement learning to scale training data
- Scores 96.2 on MATH 500 benchmark
- Matches OpenAI o1 in reasoning performance
- Avoids Monte Carlo tree search and value functions
- Outperforms GPT-4o in short-CoT reasoning
What it is
Kimi k1.5 is a multimodal large language model created by the Kimi Team. It was introduced in January 2025 through a technical paper. The system focuses on scaling reinforcement learning to enhance artificial intelligence capabilities. Unlike traditional pretraining that relies on next token prediction, this model learns to explore using rewards. This approach allows it to generate its own training data effectively. The model handles both text and visual inputs. It is designed to perform complex reasoning tasks across multiple domains.
How it came about
The Kimi Team published their findings in January 2025. The paper detailed the training practices used for this latest model. Previous work in scaling reinforcement learning had not yielded competitive results. The team aimed to address this gap by reporting their specific techniques. They focused on infrastructure optimization and multi-modal data recipes. The development process involved refining policy optimization methods. The team released updated versions of their paper through June 2025 to incorporate revisions.
Inside the model
The architecture relies on scaling reinforcement learning rather than just increasing compute for next token prediction. Key components include long context scaling and improved policy optimization. The framework is described as simplistic and effective. It avoids complex techniques like Monte Carlo tree search, value functions, or process reward models. The team employed long2short methods to enhance short chain-of-thought models. These methods use long chain-of-thought techniques to improve performance. The training infrastructure was optimized to support these RL processes.
What it can do
The model demonstrates state-of-the-art reasoning performance across various benchmarks. It scored 77.5 on AIME and 96.2 on MATH 500. On Codeforces, it reached the 94th percentile. Its MathVista score was 74.9. These results match the performance of OpenAI's o1. For short chain-of-thought reasoning, it achieved 60.8 on AIME and 94.6 on MATH 500. It scored 47.3 on LiveCodeBench. These figures outperform existing short-CoT models like GPT-4o and Claude Sonnet 3.5 by significant margins.
How to use it
Information about Kimi k1.5 is primarily available through academic channels. The Kimi Team shared details via a paper on arXiv. The document outlines training practices and technical achievements. Access appears limited to research contexts through the published literature.
Why it matters
The work suggests that scaling reinforcement learning offers a new path for AI improvement. It challenges the limitation of available training data in standard pretraining. By showing competitive results without complex auxiliary models, it provides an alternative framework. The performance gains over established models highlight the potential of RL scaling. This approach may influence future developments in reasoning-focused language models.
Questions people ask about Kimi k1.5
What makes Kimi k1.5 different from other models?
It scales reinforcement learning to generate training data through exploration. This differs from standard pretraining which relies on fixed datasets. The model uses a simple RL framework without complex additions like value functions.
How does it perform on math and coding tasks?
It achieves high scores, including 96.2 on MATH 500 and 77.5 on AIME. On Codeforces, it reaches the 94th percentile. These results match or exceed those of other leading reasoning models.
Does it use complex techniques like Monte Carlo tree search?
No. The team states their framework is simplistic and effective. It does not rely on Monte Carlo tree search, value functions, or process reward models for its performance.
Good reading and sources
There is no English Wikipedia article on Kimi k1.5 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2501.12599. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.