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.

Kimi-Audio Technical Report

We present Kimi-Audio, an open-source audio foundation model that excels in audio understanding, generation, and conversation. We detail the practices in building Kimi-Audio, including model architecture, data curation, training recipe, inference deployment, and evaluation. Specifically, we leverage a 12.5Hz audio tokenizer, design a novel LLM-based architecture with continuous features as input and discrete tokens as output, and develop a chunk-wise streaming detokenizer based on flow matching. We curate a pre-training dataset that consists of more than 13 million hours of audio data covering a wide range of modalities including speech, sound, and music, and build a pipeline to construct high-quality and diverse post-training data. Initialized from a pre-trained LLM, Kimi-Audio is continual pre-trained on both audio and text data with several carefully designed tasks, and then fine-tuned to support a diverse of audio-related tasks. Extensive evaluation shows that Kimi-Audio achieves state-of-the-art performance on a range of audio benchmarks including speech recognition, audio understanding, audio question answering, and speech conversation. We release the codes, model checkpoints, as well as the evaluation toolkits in https://github.com/MoonshotAI/Kimi-Audio.

WMKA-Net: A Weighted Multi-Kernel Attention NetworkMethod for Retinal Vessel Segmentation

We propose a novel retinal vessel segmentation network, the Weighted Multi-Kernel Attention Network (WMKA-Net), which aims to address the issues of insufficient multiscale feature capture, loss of contextual information, and noise sensitivity in retinal vessel segmentation. WMKA-Net significantly improves the segmentation performance of small vessels and low-contrast regions by integrating several innovative components, including the MultiKernelFeature Fusion Module (MKDC), the Progressive Feature Weighting Fusion Strategy (UDFF), and the Attention Mechanism Module (AttentionBlock). The MKDC module employs multiscale parallel convolutional kernels to extract vessel characteristics, thereby enhancing the ability to capture complex vascular structures. The UDFF strategy optimizes the transmission of feature information by weighted fusion of high- and low-level features. The AttentionBlock highlights key regions and suppresses noise interference through the attention mechanism. Experimental results demonstrate that WMKA-Net achieves excellent segmentation performance in multiple public datasets, particularly in segmentation of small vessels and processing of pathological regions. This work provides a robust and efficient new method for segmentation of the retinal vessel.

Muon is Scalable for LLM Training

Recently, the Muon optimizer based on matrix orthogonalization has demonstrated strong results in training small-scale language models, but the scalability to larger models has not been proven. We identify two crucial techniques for scaling up Muon: (1) adding weight decay and (2) carefully adjusting the per-parameter update scale. These techniques allow Muon to work out-of-the-box on large-scale training without the need of hyper-parameter tuning. Scaling law experiments indicate that Muon achieves $\sim\!2\times$ computational efficiency compared to AdamW with compute optimal training. Based on these improvements, we introduce Moonlight, a 3B/16B-parameter Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon. Our model improves the current Pareto frontier, achieving better performance with much fewer training FLOPs compared to prior models. We open-source our distributed Muon implementation that is memory optimal and communication efficient. We also release the pretrained, instruction-tuned, and intermediate checkpoints to support future research.

MoBA: Mixture of Block Attention for Long-Context LLMs

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

Fully Automated CTC Detection, Segmentation and Classification for Multi-Channel IF Imaging

Liquid biopsies (eg., blood draws) offer a less invasive and non-localized alternative to tissue biopsies for monitoring the progression of metastatic breast cancer (mBCa). Immunofluoresence (IF) microscopy is a tool to image and analyze millions of blood cells in a patient sample. By detecting and genetically sequencing circulating tumor cells (CTCs) in the blood, personalized treatment plans are achievable for various cancer subtypes. However, CTCs are rare (about 1 in 2M), making manual CTC detection very difficult. In addition, clinicians rely on quantitative cellular biomarkers to manually classify CTCs. This requires prior tasks of cell detection, segmentation and feature extraction. To assist clinicians, we have developed a fully automated machine learning-based production-level pipeline to efficiently detect, segment and classify CTCs in multi-channel IF images. We achieve over 99% sensitivity and 97% specificity on 9,533 cells from 15 mBCa patients. Our pipeline has been successfully deployed on real mBCa patients, reducing a patient average of 14M detected cells to only 335 CTC candidates for manual review.

Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving

Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. It features a KVCache-centric disaggregated architecture that separates the prefill and decoding clusters. It also leverages the underutilized CPU, DRAM, and SSD resources of the GPU cluster to implement a disaggregated cache of KVCache. The core of Mooncake is its KVCache-centric scheduler, which balances maximizing overall effective throughput while meeting latency-related Service Level Objectives (SLOs). Unlike traditional studies that assume all requests will be processed, Mooncake faces challenges due to highly overloaded scenarios. To mitigate these, we developed a prediction-based early rejection policy. Experiments show that Mooncake excels in long-context scenarios. Compared to the baseline method, Mooncake can achieve up to a 525% increase in throughput in certain simulated scenarios while adhering to SLOs. Under real workloads, Mooncake's innovative architecture enables Kimi to handle 75% more requests.

Predicting microRNA-disease associations from knowledge graph using tensor decomposition with relational constraints

Motivation: MiRNAs are a kind of small non-coding RNAs that are not translated into proteins, and aberrant expression of miRNAs is associated with human diseases. Since miRNAs have different roles in diseases, the miRNA-disease associations are categorized into multiple types according to their roles. Predicting miRNA-disease associations and types is critical to understand the underlying pathogenesis of human diseases from the molecular level. Results: In this paper, we formulate the problem as a link prediction in knowledge graphs. We use biomedical knowledge bases to build a knowledge graph of entities representing miRNAs and disease and multi-relations, and we propose a tensor decomposition-based model named TDRC to predict miRNA-disease associations and their types from the knowledge graph. We have experimentally evaluated our method and compared it to several baseline methods. The results demonstrate that the proposed method has high-accuracy and high-efficiency performances.