Kimi K3: Open Frontier Intelligence

We introduce Kimi K3, a 2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window. Kimi K3 is built on Kimi Delta Attention and Attention Residuals, which improve information flow across sequence length and model depth. Together with Stable LatentMoE, which effectively activates 16 of 896 routed experts per token, and refined training and data recipes, these advances yield an approximately 2.5x improvement in overall scaling efficiency over Kimi K2. Post-training highlights reinforcement learning across general, agentic, and coding domains and multiple reasoning-effort levels, enabling compositional generalization and robust long-horizon execution. At 2.8T scale, Kimi K3 is supported by infrastructure advances in multiple areas: algorithm-system co-design for KDA, perfectly balanced expert-parallel training with efficient memory management, million-token agentic RL with persistent rollout and sandbox states, and deployment innovations. Extensive evaluations show that Kimi K3 achieves frontier-level performance across long-horizon coding, agentic, knowledge, reasoning, and vision tasks. While its overall performance still trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol, Kimi K3 consistently outperforms other open and proprietary models evaluated in our suite. We release the full Kimi K3 model weights to facilitate future research and accelerate the broader deployment and adoption of frontier intelligence.

UBEP: Re-architecting Expert Parallelism Communication Library for Production Superpods

The deployment of Mixture-of-Experts (MoE) models on production high-bandwidth superpods, such as NVIDIA's NVL72/576 and Huawei's CloudMatrix384, introduces critical challenges beyond raw interconnect bandwidth. While these systems provide unified global address spaces and high-bandwidth fabrics, their full potential for sparse MoE communication is hindered by three fundamental bottlenecks: (1) Strict execution serialization imposed by coarse-grained Bulk Synchronous Parallel (BSP) orchestration of interdependent communication phases; (2) Prohibitive synchronization overhead that fails to scale alongside high interconnect bandwidth; and (3) Severe load imbalance resulting from distance-agnostic scheduling of irregular token traffic. To eliminate these bottlenecks, we introduce UBEP (Unified-Bus Expert Parallelism), a production-ready communication library that rethinks MoE's All-to-All primitives for modern superpod architectures. Through large scale experiments, UBEP reduces All-to-All latency by up to 52.4% and MoE inference Time Per Output Token (TPOT) by up to 11.1%.

EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments

Pretraining scaling laws reveal that model capability improves predictably with data and compute. But learning from real world environments after deployment remains far less understood. Analyzing roughly 38,000 hours of agent interaction with the environment across 134 real world tasks, we find, to the best of our knowledge, the first evidence that overall performance during environment learning follows a log-sigmoid scaling law with remarkably high precision, reaching R^2 = 0.998. Across model generations, we also find that agent learning speed roughly doubles every three months. This discovery stems from EdgeBench, a suite of 134 real world tasks with ultra-long horizons, spanning scientific discovery, software engineering, combinatorial optimization, professional knowledge work, formal mathematics, and interactive games. Each task sustains at least 12 hours of continuous agent operation under rich, multilevel feedback, and is built through substantial expert effort. We publicly release 51 tasks and our full evaluation framework to accelerate the study of how agents learn from real world experience.

DeepNuParc: A Novel Deep Clustering Framework for Fine-scale Parcellation of Brain Nuclei Using Diffusion MRI Tractography

Brain nuclei are clusters of anatomically distinct neurons that serve as important hubs for processing and relaying information in various neural circuits. Fine-scale parcellation of the brain nuclei is vital for a comprehensive understanding of its anatomico-functional correlations. Diffusion MRI tractography is an advanced imaging technique that can estimate the brain's white matter structural connectivity to potentially reveal the topography of the nuclei of interest for studying its subdivisions. In this work, we present a deep clustering pipeline, namely DeepNuParc, to perform automated, fine-scale parcellation of brain nuclei using diffusion MRI tractography. First, we incorporate a newly proposed deep learning approach to enable accurate segmentation of the nuclei of interest directly on the dMRI data. Next, we design a novel streamline clustering-based structural connectivity feature for a robust representation of voxels within the nuclei. Finally, we improve the popular joint dimensionality reduction and k-means clustering approach to enable nuclei parcellation at a finer scale. We demonstrate DeepNuParc on two important brain structures, i.e. the amygdala and the thalamus, that are known to have multiple anatomically and functionally distinct nuclei subdivisions. Experimental results show that DeepNuParc enables consistent parcellation of the nuclei into multiple parcels across multiple subjects and achieves good correspondence with the widely used coarse-scale atlases. Our codes are available at https://github.com/HarlandZZC/deep_nuclei_parcellation.

Differentiable Low-computation Global Correlation Loss for Monotonicity Evaluation in Quality Assessment

In this paper, we propose a global monotonicity consistency training strategy for quality assessment, which includes a differentiable, low-computation monotonicity evaluation loss function and a global perception training mechanism. Specifically, unlike conventional ranking loss and linear programming approaches that indirectly implement the Spearman rank-order correlation coefficient (SROCC) function, our method directly converts SROCC into a loss function by making the sorting operation within SROCC differentiable and functional. Furthermore, to mitigate the discrepancies between batch optimization during network training and global evaluation of SROCC, we introduce a memory bank mechanism. This mechanism stores gradient-free predicted results from previous batches and uses them in the current batch's training to prevent abrupt gradient changes. We evaluate the performance of the proposed method on both images and point clouds quality assessment tasks, demonstrating performance gains in both cases.

Learnable Scaled Gradient Descent for Guaranteed Robust Tensor PCA

Robust tensor principal component analysis (RTPCA) aims to separate the low-rank and sparse components from multi-dimensional data, making it an essential technique in the signal processing and computer vision fields. Recently emerging tensor singular value decomposition (t-SVD) has gained considerable attention for its ability to better capture the low-rank structure of tensors compared to traditional matrix SVD. However, existing methods often rely on the computationally expensive tensor nuclear norm (TNN), which limits their scalability for real-world tensors. To address this issue, we explore an efficient scaled gradient descent (SGD) approach within the t-SVD framework for the first time, and propose the RTPCA-SGD method. Theoretically, we rigorously establish the recovery guarantees of RTPCA-SGD under mild assumptions, demonstrating that with appropriate parameter selection, it achieves linear convergence to the true low-rank tensor at a constant rate, independent of the condition number. To enhance its practical applicability, we further propose a learnable self-supervised deep unfolding model, which enables effective parameter learning. Numerical experiments on both synthetic and real-world datasets demonstrate the superior performance of the proposed methods while maintaining competitive computational efficiency, especially consuming less time than RTPCA-TNN.

BCDNet: A Convolutional Neural Network For Breast Cancer Detection

Previous research has established that breast cancer is a prevalent cancer type, with Invasive Ductal Carcinoma (IDC) being the most common subtype. The incidence of this dangerous cancer continues to rise, making accurate and rapid diagnosis, particularly in the early stages, critically important. While modern Computer-Aided Diagnosis (CAD) systems can address most cases, medical professionals still face challenges in using them in the field without powerful computing resources. In this paper, we propose a novel CNN model called BCDNet, which effectively detects IDC in histopathological images with an accuracy of up to 89.5% and reduces training time effectively.

DA-Flow: Dual Attention Normalizing Flow for Skeleton-based Video Anomaly Detection

Cooperation between temporal convolutional networks (TCN) and graph convolutional networks (GCN) as a processing module has shown promising results in skeleton-based video anomaly detection (SVAD). However, to maintain a lightweight model with low computational and storage complexity, shallow GCN and TCN blocks are constrained by small receptive fields and a lack of cross-dimension interaction capture. To tackle this limitation, we propose a lightweight module called the Dual Attention Module (DAM) for capturing cross-dimension interaction relationships in spatio-temporal skeletal data. It employs the frame attention mechanism to identify the most significant frames and the skeleton attention mechanism to capture broader relationships across fixed partitions with minimal parameters and flops. Furthermore, the proposed Dual Attention Normalizing Flow (DA-Flow) integrates the DAM as a post-processing unit after GCN within the normalizing flow framework. Simulations show that the proposed model is robust against noise and negative samples. Experimental results show that DA-Flow reaches competitive or better performance than the existing state-of-the-art (SOTA) methods in terms of the micro AUC metric with the fewest number of parameters. Moreover, we found that even without training, simply using random projection without dimensionality reduction on skeleton data enables substantial anomaly detection capabilities.

Tensor Polynomial Additive Model

Additive models can be used for interpretable machine learning for their clarity and simplicity. However, In the classical models for high-order data, the vectorization operation disrupts the data structure, which may lead to degenerated accuracy and increased computational complexity. To deal with these problems, we propose the tensor polynomial addition model (TPAM). It retains the multidimensional structure information of high-order inputs with tensor representation. The model parameter compression is achieved using a hierarchical and low-order symmetric tensor approximation. In this way, complex high-order feature interactions can be captured with fewer parameters. Moreover, The TPAM preserves the inherent interpretability of additive models, facilitating transparent decision-making and the extraction of meaningful feature values. Additionally, leveraging TPAM's transparency and ability to handle higher-order features, it is used as a post-processing module for other interpretation models by introducing two variants for class activation maps. Experimental results on a series of datasets demonstrate that TPAM can enhance accuracy by up to 30\%, and compression rate by up to 5 times, while maintaining a good interpretability.

HOIN: High-Order Implicit Neural Representations

Implicit neural representations (INR) suffer from worsening spectral bias, which results in overly smooth solutions to the inverse problem. To deal with this problem, we propose a universal framework for processing inverse problems called \textbf{High-Order Implicit Neural Representations (HOIN)}. By refining the traditional cascade structure to foster high-order interactions among features, HOIN enhances the model's expressive power and mitigates spectral bias through its neural tangent kernel's (NTK) strong diagonal properties, accelerating and optimizing inverse problem resolution. By analyzing the model's expression space, high-order derivatives, and the NTK matrix, we theoretically validate the feasibility of HOIN. HOIN realizes 1 to 3 dB improvements in most inverse problems, establishing a new state-of-the-art recovery quality and training efficiency, thus providing a new general paradigm for INR and paving the way for it to solve the inverse problem.