Jianfeng Yan

Tencent Hunyuan Team, Ao Liu, Botong Zhou et al. · tencent-ai

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Jianfeng Yan, Jared Crean, Jason E. Hicken

This work focuses on multidimensional summation-by-parts (SBP) discretizations of linear elliptic operators with variable coefficients. We consider a general SBP discretization with dense simultaneous approximation terms (SATs), which serve as interior penalties to enforce boundary conditions and inter-element coupling in a weak sense. Through the analysis of adjoint consistency and stability, we present several conditions on the SAT penalties. Based on these conditions, we generalize the modified scheme of Bassi and Rebay (BR2) and the symmetric interior penalty Galerkin (SIPG) method to SBP-SAT discretizations. Numerical experiments are carried out on unstructured grids with triangular elements to verify the theoretical results.

Xingwu Sun, Yanfeng Chen, Yiqing Huang et al. · tencent-ai

In this paper, we introduce Hunyuan-Large, which is currently the largest open-source Transformer-based mixture of experts model, with a total of 389 billion parameters and 52 billion activation parameters, capable of handling up to 256K tokens. We conduct a thorough evaluation of Hunyuan-Large's superior performance across various benchmarks including language understanding and generation, logical reasoning, mathematical problem-solving, coding, long-context, and aggregated tasks, where it outperforms LLama3.1-70B and exhibits comparable performance when compared to the significantly larger LLama3.1-405B model. Key practice of Hunyuan-Large include large-scale synthetic data that is orders larger than in previous literature, a mixed expert routing strategy, a key-value cache compression technique, and an expert-specific learning rate strategy. Additionally, we also investigate the scaling laws and learning rate schedule of mixture of experts models, providing valuable insights and guidances for future model development and optimization. The code and checkpoints of Hunyuan-Large are released to facilitate future innovations and applications. Codes: https://github.com/Tencent/Hunyuan-Large Models: https://huggingface.co/tencent/Tencent-Hunyuan-Large

Jianfeng Yan, Zizhuang Wei, Hongwei Yi et al.

In this paper, we propose an efficient and effective dense hybrid recurrent multi-view stereo net with dynamic consistency checking, namely $D^{2}$HC-RMVSNet, for accurate dense point cloud reconstruction. Our novel hybrid recurrent multi-view stereo net consists of two core modules: 1) a light DRENet (Dense Reception Expanded) module to extract dense feature maps of original size with multi-scale context information, 2) a HU-LSTM (Hybrid U-LSTM) to regularize 3D matching volume into predicted depth map, which efficiently aggregates different scale information by coupling LSTM and U-Net architecture. To further improve the accuracy and completeness of reconstructed point clouds, we leverage a dynamic consistency checking strategy instead of prefixed parameters and strategies widely adopted in existing methods for dense point cloud reconstruction. In doing so, we dynamically aggregate geometric consistency matching error among all the views. Our method ranks \textbf{$1^{st}$} on the complex outdoor \textsl{Tanks and Temples} benchmark over all the methods. Extensive experiments on the in-door DTU dataset show our method exhibits competitive performance to the state-of-the-art method while dramatically reduces memory consumption, which costs only $19.4\%$ of R-MVSNet memory consumption. The codebase is available at \hyperlink{https://github.com/yhw-yhw/D2HC-RMVSNet}{https://github.com/yhw-yhw/D2HC-RMVSNet}.

Siheng Li, Kejiao Li, Zenan Xu et al.

The growing disparity between the exponential scaling of computational resources and the finite growth of high-quality text data now constrains conventional scaling approaches for large language models (LLMs). To address this challenge, we introduce Reinforcement Learning on Pre-Training data (RLPT), a new training-time scaling paradigm for optimizing LLMs. In contrast to prior approaches that scale training primarily through supervised learning, RLPT enables the policy to autonomously explore meaningful trajectories to learn from pre-training data and improve its capability through reinforcement learning (RL). While existing RL strategies such as reinforcement learning from human feedback (RLHF) and reinforcement learning with verifiable rewards (RLVR) rely on human annotation for reward construction, RLPT eliminates this dependency by deriving reward signals directly from pre-training data. Specifically, it adopts a next-segment reasoning objective, rewarding the policy for accurately predicting subsequent text segments conditioned on the preceding context. This formulation allows RL to be scaled on pre-training data, encouraging the exploration of richer trajectories across broader contexts and thereby fostering more generalizable reasoning skills. Extensive experiments on both general-domain and mathematical reasoning benchmarks across multiple models validate the effectiveness of RLPT. For example, when applied to Qwen3-4B-Base, RLPT yields absolute improvements of $3.0$, $5.1$, $8.1$, $6.0$, $6.6$, and $5.3$ on MMLU, MMLU-Pro, GPQA-Diamond, KOR-Bench, AIME24, and AIME25, respectively. The results further demonstrate favorable scaling behavior, suggesting strong potential for continued gains with more compute. In addition, RLPT provides a solid foundation, extending the reasoning boundaries of LLMs and enhancing RLVR performance.

Yuyuan Zeng, Yufei Huang, Can Xu et al.

Zero Reinforcement Learning (Zero-RL) has proven to be an effective approach for enhancing the reasoning capabilities of large language models (LLMs) by directly applying reinforcement learning with verifiable rewards on pretrained models, without the need for a supervised fine-tuning phase. However, current research on zero-RL primarily focuses on domains with easily verifiable reward signals, such as mathematics, programming, and other reasoning tasks. The challenge of eliciting reasoning abilities in more diverse scenarios, where verification is not straightforward, remains underexplored. To address this gap, we propose a novel zero-RL paradigm designed to improve a model's reasoning ability across both verifiable and non-verifiable domains. By combining verifiable rewards with a generative reward model, we conduct multi-task zero-RL training across both domains, facilitating the transfer of reasoning capabilities between them. Furthermore, to mitigate reward hacking in the generative reward model, we design a smooth length penalty that encourages the generation of more comprehensive thinking tokens in general domains. Experimental results on Qwen3-8B-Base and Qwen3-14B-Base demonstrate that our approach achieves superior reasoning performance, not only on tasks requiring extensive reasoning but also on more general tasks.

Han Xiao, Feng Wang, Jianfeng Yan et al.

There are three modalities in the reading comprehension setting: question, answer and context. The task of question answering or question generation aims to infer an answer or a question when given the counterpart based on context. We present a novel two-way neural sequence transduction model that connects three modalities, allowing it to learn two tasks simultaneously and mutually benefit one another. During training, the model receives question-context-answer triplets as input and captures the cross-modal interaction via a hierarchical attention process. Unlike previous joint learning paradigms that leverage the duality of question generation and question answering at data level, we solve such dual tasks at the architecture level by mirroring the network structure and partially sharing components at different layers. This enables the knowledge to be transferred from one task to another, helping the model to find a general representation for each modality. The evaluation on four public datasets shows that our dual-learning model outperforms the mono-learning counterpart as well as the state-of-the-art joint models on both question answering and question generation tasks.