MoonshotAI has announced the release of Kimi k1.5, a cutting-edge multi-modal large language model (LLM) designed to advance reasoning capabilities across text and vision modalities. The model, launched on January 19, 2025, represents a significant step forward in reinforcement learning (RL) applications for LLMs, achieving state-of-the-art (SOTA) performance in various benchmarks such as AIME, MATH500, and Codeforces.
Key Features of Kimi k1.5
- Performance:
- Outperforms GPT-4o and Claude Sonnet 3.5 in short-chain-of-thought (short-CoT) reasoning tasks by up to 550%.
- Matches OpenAI's o1 model in long-chain-of-thought (long-CoT) reasoning across multiple domains.
- Benchmark scores include 77.5 on AIME, 96.2 on MATH500, and the 94th percentile on Codeforces.
- Technical Advancements:
- Long Context Scaling: Supports context windows of up to 128k tokens, enabling improved reasoning over extended inputs.
- Simplified RL Framework: Utilizes efficient techniques like partial rollouts and avoids complex methods such as Monte Carlo tree search.
- Improved Policy Optimization: Incorporates advanced sampling strategies and optimization techniques for robust training.
- Multi-Modality: Jointly trained on text and vision data, allowing seamless reasoning across both formats.
- Applications:
- Excels in mathematics, coding, and visual reasoning tasks.
- Designed to tackle complex problems requiring deep reasoning and planning.
- API Access:
- A preview version is available through the Kimi OpenPlatform for testing via API.
The Kimi k1.5 model underscores MoonshotAI's commitment to pushing the boundaries of AI reasoning capabilities. By leveraging reinforcement learning to scale training data through exploration with rewards, the model addresses limitations in traditional pretraining methods reliant on static datasets. Its ability to handle long contexts and multi-modal inputs positions it as a competitive alternative to leading models like OpenAI's o1.

About MoonshotAI
MoonshotAI is an emerging player in the AI landscape, focusing on developing advanced multi-modal models. With Kimi k1.5, the company aims to establish itself as a leader in AI-driven reasoning solutions, directly competing with established names like OpenAI and Anthropic.
The Kimi k1.5 model is set to be made publicly available soon through its official platform at kimi.ai, marking another milestone in the evolution of AI technology for diverse applications.