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.

AnimeColor: Reference-based Animation Colorization with Diffusion Transformers

Animation colorization plays a vital role in animation production, yet existing methods struggle to achieve color accuracy and temporal consistency. To address these challenges, we propose \textbf{AnimeColor}, a novel reference-based animation colorization framework leveraging Diffusion Transformers (DiT). Our approach integrates sketch sequences into a DiT-based video diffusion model, enabling sketch-controlled animation generation. We introduce two key components: a High-level Color Extractor (HCE) to capture semantic color information and a Low-level Color Guider (LCG) to extract fine-grained color details from reference images. These components work synergistically to guide the video diffusion process. Additionally, we employ a multi-stage training strategy to maximize the utilization of reference image color information. Extensive experiments demonstrate that AnimeColor outperforms existing methods in color accuracy, sketch alignment, temporal consistency, and visual quality. Our framework not only advances the state of the art in animation colorization but also provides a practical solution for industrial applications. The code will be made publicly available at \href{https://github.com/IamCreateAI/AnimeColor}{https://github.com/IamCreateAI/AnimeColor}.

StepHint: Multi-level Stepwise Hints Enhance Reinforcement Learning to Reason

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for improving the complex reasoning abilities of large language models (LLMs). However, current RLVR methods face two significant challenges: the near-miss reward problem, where a small mistake can invalidate an otherwise correct reasoning process, greatly hindering training efficiency; and exploration stagnation, where models tend to focus on solutions within their ``comfort zone,'' lacking the motivation to explore potentially more effective alternatives. To address these challenges, we propose StepHint, a novel RLVR algorithm that utilizes multi-level stepwise hints to help models explore the solution space more effectively. StepHint generates valid reasoning chains from stronger models and partitions these chains into reasoning steps using our proposed adaptive partitioning method. The initial few steps are used as hints, and simultaneously, multiple-level hints (each comprising a different number of steps) are provided to the model. This approach directs the model's exploration toward a promising solution subspace while preserving its flexibility for independent exploration. By providing hints, StepHint mitigates the near-miss reward problem, thereby improving training efficiency. Additionally, the external reasoning pathways help the model develop better reasoning abilities, enabling it to move beyond its ``comfort zone'' and mitigate exploration stagnation. StepHint outperforms competitive RLVR enhancement methods across six mathematical benchmarks, while also demonstrating superior generalization and excelling over baselines on out-of-domain benchmarks.

Constellation as a Service: Tailored Connectivity Management in Direct-Satellite-to-Device Networks

Direct-satellite-to-device (DS2D) communication is emerging as a promising solution for global mobile service extension, leveraging the deployment of satellite constellations. However, the challenge of managing DS2D connectivity for multi-constellations becomes outstanding, including high interference and frequent handovers caused by multi-coverage overlap and rapid satellite movement. Moreover, existing approaches primarily operate within single-constellation shell, which inherently limits the ability to exploit the vast potential of multi-constellation connectivity provision, resulting in suboptimal DS2D service performances. To address these challenges, this article proposes a Constellation as a Service (CaaS) framework, which treats the entire multi-constellation infrastructure as a shared resource pool and dynamically forms optimal sub-constellations (SCs) for each DS2D service region. The formation of each SC integrates satellites from various orbits to provide tailored connectivity based on user demands, guided by two innovative strategies: predictive satellite beamforming using generative artificial intelligence (GenAI) and pre-configured handover path for efficient satellite access and mobility management. Simulation results demonstrate that CaaS significantly improves satellite service rates while reducing handover overhead, making it an efficient and continuable solution for managing DS2D connectivity in multi-constellation environments.

MagCache: Fast Video Generation with Magnitude-Aware Cache

Existing acceleration techniques for video diffusion models often rely on uniform heuristics or time-embedding variants to skip timesteps and reuse cached features. These approaches typically require extensive calibration with curated prompts and risk inconsistent outputs due to prompt-specific overfitting. In this paper, we introduce a novel and robust discovery: a unified magnitude law observed across different models and prompts. Specifically, the magnitude ratio of successive residual outputs decreases monotonically and steadily in most timesteps while rapidly in the last several steps. Leveraging this insight, we introduce a Magnitude-aware Cache (MagCache) that adaptively skips unimportant timesteps using an error modeling mechanism and adaptive caching strategy. Unlike existing methods requiring dozens of curated samples for calibration, MagCache only requires a single sample for calibration. Experimental results show that MagCache achieves 2.1x and 2.68x speedups on Open-Sora and Wan 2.1, respectively, while preserving superior visual fidelity. It significantly outperforms existing methods in LPIPS, SSIM, and PSNR, under comparable computational budgets.

MiniCPM4: Ultra-Efficient LLMs on End Devices

This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and UltraChat v2, a comprehensive supervised fine-tuning dataset. These datasets enable satisfactory model performance to be achieved using just 8 trillion training tokens. Regarding training algorithms, we propose ModelTunnel v2 for efficient pre-training strategy search, and improve existing post-training methods by introducing chunk-wise rollout for load-balanced reinforcement learning and data-efficient tenary LLM, BitCPM. Regarding inference systems, we propose CPM.cu that integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding. To meet diverse on-device requirements, MiniCPM4 is available in two versions, with 0.5B and 8B parameters, respectively. Sufficient evaluation results show that MiniCPM4 outperforms open-source models of similar size across multiple benchmarks, highlighting both its efficiency and effectiveness. Notably, MiniCPM4-8B demonstrates significant speed improvements over Qwen3-8B when processing long sequences. Through further adaptation, MiniCPM4 successfully powers diverse applications, including trustworthy survey generation and tool use with model context protocol, clearly showcasing its broad usability.

CLIP-aware Domain-Adaptive Super-Resolution

This work introduces CLIP-aware Domain-Adaptive Super-Resolution (CDASR), a novel framework that addresses the critical challenge of domain generalization in single image super-resolution. By leveraging the semantic capabilities of CLIP (Contrastive Language-Image Pre-training), CDASR achieves unprecedented performance across diverse domains and extreme scaling factors. The proposed method integrates CLIP-guided feature alignment mechanism with a meta-learning inspired few-shot adaptation strategy, enabling efficient knowledge transfer and rapid adaptation to target domains. A custom domain-adaptive module processes CLIP features alongside super-resolution features through a multi-stage transformation process, including CLIP feature processing, spatial feature generation, and feature fusion. This intricate process ensures effective incorporation of semantic information into the super-resolution pipeline. Additionally, CDASR employs a multi-component loss function that combines pixel-wise reconstruction, perceptual similarity, and semantic consistency. Extensive experiments on benchmark datasets demonstrate CDASR's superiority, particularly in challenging scenarios. On the Urban100 dataset at $\times$8 scaling, CDASR achieves a significant PSNR gain of 0.15dB over existing methods, with even larger improvements of up to 0.30dB observed at $\times$16 scaling.

Differentiable NMS via Sinkhorn Matching for End-to-End Fabric Defect Detection

Fabric defect detection confronts two fundamental challenges. First, conventional non-maximum suppression disrupts gradient flow, which hinders genuine end-to-end learning. Second, acquiring pixel-level annotations at industrial scale is prohibitively costly. Addressing these limitations, we propose a differentiable NMS framework for fabric defect detection that achieves superior localization precision through end-to-end optimization. We reformulate NMS as a differentiable bipartite matching problem solved through the Sinkhorn-Knopp algorithm, maintaining uninterrupted gradient flow throughout the network. This approach specifically targets the irregular morphologies and ambiguous boundaries of fabric defects by integrating proposal quality, feature similarity, and spatial relationships. Our entropy-constrained mask refinement mechanism further enhances localization precision through principled uncertainty modeling. Extensive experiments on the Tianchi fabric defect dataset demonstrate significant performance improvements over existing methods while maintaining real-time speeds suitable for industrial deployment. The framework exhibits remarkable adaptability across different architectures and generalizes effectively to general object detection tasks.

ReaderLM-v2: Small Language Model for HTML to Markdown and JSON

We present ReaderLM-v2, a compact 1.5 billion parameter language model designed for efficient web content extraction. Our model processes documents up to 512K tokens, transforming messy HTML into clean Markdown or JSON formats with high accuracy -- making it an ideal tool for grounding large language models. The model's effectiveness results from two key innovations: (1) a three-stage data synthesis pipeline that generates high quality, diverse training data by iteratively drafting, refining, and critiquing web content extraction; and (2) a unified training framework combining continuous pre-training with multi-objective optimization. Intensive evaluation demonstrates that ReaderLM-v2 outperforms GPT-4o-2024-08-06 and other larger models by 15-20\% on carefully curated benchmarks, particularly excelling at documents exceeding 100K tokens, while maintaining significantly lower computational requirements.

Scaling Laws in Patchification: An Image Is Worth 50,176 Tokens And More

Since the introduction of Vision Transformer (ViT), patchification has long been regarded as a de facto image tokenization approach for plain visual architectures. By compressing the spatial size of images, this approach can effectively shorten the token sequence and reduce the computational cost of ViT-like plain architectures. In this work, we aim to thoroughly examine the information loss caused by this patchification-based compressive encoding paradigm and how it affects visual understanding. We conduct extensive patch size scaling experiments and excitedly observe an intriguing scaling law in patchification: the models can consistently benefit from decreased patch sizes and attain improved predictive performance, until it reaches the minimum patch size of 1x1, i.e., pixel tokenization. This conclusion is broadly applicable across different vision tasks, various input scales, and diverse architectures such as ViT and the recent Mamba models. Moreover, as a by-product, we discover that with smaller patches, task-specific decoder heads become less critical for dense prediction. In the experiments, we successfully scale up the visual sequence to an exceptional length of 50,176 tokens, achieving a competitive test accuracy of 84.6% with a base-sized model on the ImageNet-1k benchmark. We hope this study can provide insights and theoretical foundations for future works of building non-compressive vision models. Code is available at https://github.com/wangf3014/Patch_Scaling.