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.

MBD: A Model-Based Debiasing Framework Across User, Content, and Model Dimensions

Modern recommendation systems rank candidates by aggregating multiple behavioral signals through a value model. However, many commonly used signals are inherently affected by heterogeneous biases. For example, watch time naturally favors long-form content, loop rate favors short - form content, and comment probability favors videos over images. Such biases introduce two critical issues: (1) value model scores may be systematically misaligned with users' relative preferences - for instance, a seemingly low absolute like probability may represent exceptionally strong interest for a user who rarely engages; and (2) changes in value modeling rules can trigger abrupt and undesirable ecosystem shifts. In this work, we ask a fundamental question: can biased behavioral signals be systematically transformed into unbiased signals, under a user - defined notion of ``unbiasedness'', that are both personalized and adaptive? We propose a general, model-based debiasing (MBD) framework that addresses this challenge by augmenting it with distributional modeling. By conditioning on a flexible subset of features (partial feature set), we explicitly estimate the contextual mean and variance of the engagement distribution for arbitrary cohorts (e.g., specific video lengths or user regions) directly alongside the main prediction. This integration allows the framework to convert biased raw signals into unbiased representations, enabling the construction of higher-level, calibrated signals (such as percentiles or z - scores) suitable for the value model. Importantly, the definition of unbiasedness is flexible and controllable, allowing the system to adapt to different personalization objectives and modeling preferences. Crucially, this is implemented as a lightweight, built-in branch of the existing MTML ranking model, requiring no separate serving infrastructure.

Trustworthy Evaluation of Robotic Manipulation: A New Benchmark and AutoEval Methods

Driven by the rapid evolution of Vision-Action and Vision-Language-Action models, imitation learning has significantly advanced robotic manipulation capabilities. However, evaluation methodologies have lagged behind, hindering the establishment of Trustworthy Evaluation for these behaviors. Current paradigms rely on binary success rates, failing to address the critical dimensions of trust: Source Authenticity (i.e., distinguishing genuine policy behaviors from human teleoperation) and Execution Quality (e.g., smoothness and safety). To bridge these gaps, we propose a solution that combines the Eval-Actions benchmark and the AutoEval architecture. First, we construct the Eval-Actions benchmark to support trustworthiness analysis. Distinct from existing datasets restricted to successful human demonstrations, Eval-Actions integrates VA and VLA policy execution trajectories alongside human teleoperation data, explicitly including failure scenarios. This dataset is structured around three core supervision signals: Expert Grading (EG), Rank-Guided preferences (RG), and Chain-of-Thought (CoT). Building on this, we propose the AutoEval architecture: AutoEval leverages Spatio-Temporal Aggregation for semantic assessment, augmented by an auxiliary Kinematic Calibration Signal to refine motion smoothness; AutoEval Plus (AutoEval-P) incorporates the Group Relative Policy Optimization (GRPO) paradigm to enhance logical reasoning capabilities. Experiments show AutoEval achieves Spearman's Rank Correlation Coefficients (SRCC) of 0.81 and 0.84 under the EG and RG protocols, respectively. Crucially, the framework possesses robust source discrimination capabilities, distinguishing between policy-generated and teleoperated videos with 99.6% accuracy, thereby establishing a rigorous standard for trustworthy robotic evaluation. Our project and code are available at https://term-bench.github.io/.