Kimi K2: Open Agentic Intelligence

We introduce Kimi K2, a Mixture-of-Experts (MoE) large language model with 32 billion activated parameters and 1 trillion total parameters. We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability while enjoying the advanced token efficiency of Muon. Based on MuonClip, K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, with strengths in agentic capabilities. Notably, K2 obtains 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual -- surpassing most open and closed-sourced baselines in non-thinking settings. It also exhibits strong capabilities in coding, mathematics, and reasoning tasks, with a score of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, and 27.1 on OJBench, all without extended thinking. These results position Kimi K2 as one of the most capable open-source large language models to date, particularly in software engineering and agentic tasks. We release our base and post-trained model checkpoints to facilitate future research and applications of agentic intelligence.

MetaKRec: Collaborative Meta-Knowledge Enhanced Recommender System

Knowledge graph (KG) enhanced recommendation has demonstrated improved performance in the recommendation system (RecSys) and attracted considerable research interest. Recently the literature has adopted neural graph networks (GNNs) on the collaborative knowledge graph and built an end-to-end KG-enhanced RecSys. However, the majority of these approaches have three limitations: (1) treat the collaborative knowledge graph as a homogeneous graph and overlook the highly heterogeneous relationships among items, (2) lack of design to explicitly leverage the rich side information, and (3) overlook the rich knowledge in user preference. To fill this gap, in this paper, we explore the rich, heterogeneous relationship among items and propose a new KG-enhanced recommendation model called Collaborative Meta-Knowledge Enhanced Recommender System (MetaKRec). In particular, we focus on modeling the rich, heterogeneous semantic relationships among items and construct several collaborative Meta-KGs to explicitly depict the relatedness of the items under the guidance of meta-knowledge. In addition to the knowledge obtained from KG, we leverage user knowledge that extracts from user preference to construct the Meta-KGs. The constructed Meta-KGs can capture the knowledge from both the knowledge graph and user preference. Furthermore. we utilize a light convolution encoder to recursively integrate the item relationship in each collaborative Meta-KGs. This scheme allows us to explicitly gather the heterogeneous semantic relationships among items and encode them into the representations of items. In addition, we propose channel attention to fuse the item and user representations from different Meta-KGs. Extensive experiments are conducted on four real-world benchmark datasets, demonstrating significant gains over the state-of-the-art baselines on both regular and cold-start recommendation settings.