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.

Compositional Function Networks: A High-Performance Alternative to Deep Neural Networks with Built-in Interpretability

Deep Neural Networks (DNNs) deliver impressive performance but their black-box nature limits deployment in high-stakes domains requiring transparency. We introduce Compositional Function Networks (CFNs), a novel framework that builds inherently interpretable models by composing elementary mathematical functions with clear semantics. Unlike existing interpretable approaches that are limited to simple additive structures, CFNs support diverse compositional patterns -- sequential, parallel, and conditional -- enabling complex feature interactions while maintaining transparency. A key innovation is that CFNs are fully differentiable, allowing efficient training through standard gradient descent. We demonstrate CFNs' versatility across multiple domains, from symbolic regression to image classification with deep hierarchical networks. Our empirical evaluation shows CFNs achieve competitive performance against black-box models (96.24% accuracy on CIFAR-10) while outperforming state-of-the-art interpretable models like Explainable Boosting Machines. By combining the hierarchical expressiveness and efficient training of deep learning with the intrinsic interpretability of well-defined mathematical functions, CFNs offer a powerful framework for applications where both performance and accountability are paramount.

ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving

Although end-to-end autonomous driving has made remarkable progress, its performance degrades significantly in rare and long-tail scenarios. Recent approaches attempt to address this challenge by leveraging the rich world knowledge of Vision-Language Models (VLMs), but these methods suffer from several limitations: (1) a significant domain gap between the pre-training data of VLMs and real-world driving data, (2) a dimensionality mismatch between the discrete language space and the continuous action space, and (3) imitation learning tends to capture the average behavior present in the dataset, which may be suboptimal even dangerous. In this paper, we propose ReCogDrive, an autonomous driving system that integrates VLMs with diffusion planner, which adopts a three-stage paradigm for training. In the first stage, we use a large-scale driving question-answering datasets to train the VLMs, mitigating the domain discrepancy between generic content and real-world driving scenarios. In the second stage, we employ a diffusion-based planner to perform imitation learning, mapping representations from the latent language space to continuous driving actions. Finally, we fine-tune the diffusion planner using reinforcement learning with NAVSIM non-reactive simulator, enabling the model to generate safer, more human-like driving trajectories. We evaluate our approach on the planning-oriented NAVSIM benchmark, achieving a PDMS of 89.6 and setting a new state-of-the-art that surpasses the previous vision-only SOTA by 5.6 PDMS.

Learning Cocoercive Conservative Denoisers via Helmholtz Decomposition for Poisson Inverse Problems

Plug-and-play (PnP) methods with deep denoisers have shown impressive results in imaging problems. They typically require strong convexity or smoothness of the fidelity term and a (residual) non-expansive denoiser for convergence. These assumptions, however, are violated in Poisson inverse problems, and non-expansiveness can hinder denoising performance. To address these challenges, we propose a cocoercive conservative (CoCo) denoiser, which may be (residual) expansive, leading to improved denoising. By leveraging the generalized Helmholtz decomposition, we introduce a novel training strategy that combines Hamiltonian regularization to promote conservativeness and spectral regularization to ensure cocoerciveness. We prove that CoCo denoiser is a proximal operator of a weakly convex function, enabling a restoration model with an implicit weakly convex prior. The global convergence of PnP methods to a stationary point of this restoration model is established. Extensive experimental results demonstrate that our approach outperforms closely related methods in both visual quality and quantitative metrics.

Kimi-VL Technical Report

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), this model exhibits strong long-horizon reasoning capabilities. It achieves scores of 61.7 on MMMU, 36.8 on MathVision, and 71.3 on MathVista while maintaining the compact 2.8B activated LLM parameters, setting a new standard for efficient multimodal thinking models. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

Sequential-NIAH: A Needle-In-A-Haystack Benchmark for Extracting Sequential Needles from Long Contexts

Evaluating the ability of large language models (LLMs) to handle extended contexts is critical, particularly for retrieving information relevant to specific queries embedded within lengthy inputs. We introduce Sequential-NIAH, a benchmark specifically designed to evaluate the capability of LLMs to extract sequential information items (known as needles) from long contexts. The benchmark comprises three types of needle generation pipelines: synthetic, real, and open-domain QA. It includes contexts ranging from 8K to 128K tokens in length, with a dataset of 14,000 samples (2,000 reserved for testing). To facilitate evaluation on this benchmark, we trained a synthetic data-driven evaluation model capable of evaluating answer correctness based on chronological or logical order, achieving an accuracy of 99.49% on synthetic test data. We conducted experiments on six well-known LLMs, revealing that even the best-performing model achieved a maximum accuracy of only 63.15%. Further analysis highlights the growing challenges posed by increasing context lengths and the number of needles, underscoring substantial room for improvement. Additionally, noise robustness experiments validate the reliability of the benchmark, making Sequential-NIAH an important reference for advancing research on long text extraction capabilities of LLMs.

FactGuard: Leveraging Multi-Agent Systems to Generate Answerable and Unanswerable Questions for Enhanced Long-Context LLM Extraction

Extractive reading comprehension systems are designed to locate the correct answer to a question within a given text. However, a persistent challenge lies in ensuring these models maintain high accuracy in answering questions while reliably recognizing unanswerable queries. Despite significant advances in large language models (LLMs) for reading comprehension, this issue remains critical, particularly as the length of supported contexts continues to expand. To address this challenge, we propose an innovative data augmentation methodology grounded in a multi-agent collaborative framework. Unlike traditional methods, such as the costly human annotation process required for datasets like SQuAD 2.0, our method autonomously generates evidence-based question-answer pairs and systematically constructs unanswerable questions. Using this methodology, we developed the FactGuard-Bench dataset, which comprises 25,220 examples of both answerable and unanswerable question scenarios, with context lengths ranging from 8K to 128K. Experimental evaluations conducted on seven popular LLMs reveal that even the most advanced models achieve only 61.79% overall accuracy. Furthermore, we emphasize the importance of a model's ability to reason about unanswerable questions to avoid generating plausible but incorrect answers. By implementing efficient data selection and generation within the multi-agent collaborative framework, our method significantly reduces the traditionally high costs associated with manual annotation and provides valuable insights for the training and optimization of LLMs.

TiGDistill-BEV: Multi-view BEV 3D Object Detection via Target Inner-Geometry Learning Distillation

Accurate multi-view 3D object detection is essential for applications such as autonomous driving. Researchers have consistently aimed to leverage LiDAR's precise spatial information to enhance camera-based detectors through methods like depth supervision and bird-eye-view (BEV) feature distillation. However, existing approaches often face challenges due to the inherent differences between LiDAR and camera data representations. In this paper, we introduce the TiGDistill-BEV, a novel approach that effectively bridges this gap by leveraging the strengths of both sensors. Our method distills knowledge from diverse modalities(e.g., LiDAR) as the teacher model to a camera-based student detector, utilizing the Target Inner-Geometry learning scheme to enhance camera-based BEV detectors through both depth and BEV features by leveraging diverse modalities. Specially, we propose two key modules: an inner-depth supervision module to learn the low-level relative depth relations within objects which equips detectors with a deeper understanding of object-level spatial structures, and an inner-feature BEV distillation module to transfer high-level semantics of different key points within foreground targets. To further alleviate the domain gap, we incorporate both inter-channel and inter-keypoint distillation to model feature similarity. Extensive experiments on the nuScenes benchmark demonstrate that TiGDistill-BEV significantly boosts camera-based only detectors achieving a state-of-the-art with 62.8% NDS and surpassing previous methods by a significant margin. The codes is available at: https://github.com/Public-BOTs/TiGDistill-BEV.git.

Acquire Precise and Comparable Fundus Image Quality Score: FTHNet and FQS Dataset

The retinal fundus images are utilized extensively in the diagnosis, and their quality can directly affect the diagnosis results. However, due to the insufficient dataset and algorithm application, current fundus image quality assessment (FIQA) methods are not powerful enough to meet ophthalmologists` demands. In this paper, we address the limitations of datasets and algorithms in FIQA. First, we establish a new FIQA dataset, Fundus Quality Score(FQS), which includes 2246 fundus images with two labels: a continuous Mean Opinion Score varying from 0 to 100 and a three-level quality label. Then, we propose a FIQA Transformer-based Hypernetwork (FTHNet) to solve these tasks with regression results rather than classification results in conventional FIQA works. The FTHNet is optimized for the FIQA tasks with extensive experiments. Results on our FQS dataset show that the FTHNet can give quality scores for fundus images with PLCC of 0.9423 and SRCC of 0.9488, significantly outperforming other methods with fewer parameters and less computation complexity.We successfully build a dataset and model addressing the problems of current FIQA methods. Furthermore, the model deployment experiments demonstrate its potential in automatic medical image quality control. All experiments are carried out with 10-fold cross-validation to ensure the significance of the results.

GaussianPretrain: A Simple Unified 3D Gaussian Representation for Visual Pre-training in Autonomous Driving

Self-supervised learning has made substantial strides in image processing, while visual pre-training for autonomous driving is still in its infancy. Existing methods often focus on learning geometric scene information while neglecting texture or treating both aspects separately, hindering comprehensive scene understanding. In this context, we are excited to introduce GaussianPretrain, a novel pre-training paradigm that achieves a holistic understanding of the scene by uniformly integrating geometric and texture representations. Conceptualizing 3D Gaussian anchors as volumetric LiDAR points, our method learns a deepened understanding of scenes to enhance pre-training performance with detailed spatial structure and texture, achieving that 40.6% faster than NeRF-based method UniPAD with 70% GPU memory only. We demonstrate the effectiveness of GaussianPretrain across multiple 3D perception tasks, showing significant performance improvements, such as a 7.05% increase in NDS for 3D object detection, boosts mAP by 1.9% in HD map construction and 0.8% improvement on Occupancy prediction. These significant gains highlight GaussianPretrain's theoretical innovation and strong practical potential, promoting visual pre-training development for autonomous driving. Source code will be available at https://github.com/Public-BOTs/GaussianPretrain