Shuaipeng Li

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.

Mao Zheng, Zheng Li, Tao Chen et al.

Hy-MT2 is a family of fast-thinking multilingual translation models designed for complex real-world scenarios. It includes three model sizes: 1.8B, 7B, and 30B-A3B (MoE), all of which support translation among 33 languages and effectively follow translation instructions in multiple languages. For on-device deployment, with AngelSlim 1.25-bit extreme quantization, the 1.8B model requires only 440 MB of storage and improves inference speed by 1.5x. Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model also surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall.

An Wang, Xingwu Sun, Ruobing Xie et al.

Mixture of Experts (MoE) offers remarkable performance and computational efficiency by selectively activating subsets of model parameters. Traditionally, MoE models use homogeneous experts, each with identical capacity. However, varying complexity in input data necessitates experts with diverse capabilities, while homogeneous MoE hinders effective expert specialization and efficient parameter utilization. In this study, we propose a novel Heterogeneous Mixture of Experts (HMoE), where experts differ in size and thus possess diverse capacities. This heterogeneity allows for more specialized experts to handle varying token complexities more effectively. To address the imbalance in expert activation, we propose a novel training objective that encourages the frequent activation of smaller experts, enhancing computational efficiency and parameter utilization. Extensive experiments demonstrate that HMoE achieves lower loss with fewer activated parameters and outperforms conventional homogeneous MoE models on various pre-training evaluation benchmarks. Codes will be released upon acceptance.

Pinxue Zhao, Hailin Zhang, Fangcheng Fu et al.

Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing memory reuse across transformer layers. Empirical results demonstrate that MEMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.

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

Guoliang Zhao, Ruobing Xie, An Wang et al.

As Large Language Models (LLMs) scale up, inference efficiency becomes a critical bottleneck. Multi-Token Prediction (MTP) could accelerate LLM inference by predicting multiple future tokens in parallel. However, existing MTP approaches still face two challenges: limited acceptance rates of MTP heads, and difficulties in jointly training multiple MTP heads. Therefore, we propose MTP-D, a simple yet effective self-distillation method with minimal additional training cost, which boosts MTP head acceptance rates (+7.5\%) while maximumly preserving main-head performance. We also introduce a looped extension strategy for MTP-D, enabling effective and economical MTP head extension and further significant inference speedup to 1-head MTP (+220.4\%). Moreover, we systematically explore and validate key insights on the distillation strategies and the potential scalability of MTP through extensive experiments on seven benchmarks. These results demonstrate that our MTP-D and looped extension strategy effectively enhance MTP-head performance and inference efficiency, facilitating the practical usage of MTP in LLMs.

Shuaipeng Li, Penghao Zhao, Hailin Zhang et al.

In current deep learning tasks, Adam style optimizers such as Adam, Adagrad, RMSProp, Adafactor, and Lion have been widely used as alternatives to SGD style optimizers. These optimizers typically update model parameters using the sign of gradients, resulting in more stable convergence curves. The learning rate and the batch size are the most critical hyperparameters for optimizers, which require careful tuning to enable effective convergence. Previous research has shown that the optimal learning rate increases linearly or follows similar rules with batch size for SGD style optimizers. However, this conclusion is not applicable to Adam style optimizers. In this paper, we elucidate the connection between optimal learning rates and batch sizes for Adam style optimizers through both theoretical analysis and extensive experiments. First, we raise the scaling law between batch sizes and optimal learning rates in the sign of gradient case, in which we prove that the optimal learning rate first rises and then falls as the batch size increases. Moreover, the peak value of the surge will gradually move toward the larger batch size as training progresses. Second, we conducted experiments on various CV and NLP tasks and verified the correctness of the scaling law.

Yixing Li, Ruobing Xie, Zhen Yang et al.

Transformers are the cornerstone of modern large language models, but their quadratic computational complexity limits efficiency in long-sequence processing. Recent advancements in Mamba, a state space model (SSM) with linear complexity, offer promising efficiency gains but suffer from unstable contextual learning and multitask generalization. This paper proposes TransMamba, a novel framework that unifies Transformer and Mamba through shared parameter matrices (e.g., QKV and CBx), and thus could dynamically switch between attention and SSM mechanisms at different token lengths and layers. We design the Memory converter to bridge Transformer and Mamba by converting attention outputs into SSM-compatible states, ensuring seamless information flow at TransPoints where the transformation happens. The TransPoint scheduling is also thoroughly explored for further improvements. We conducted extensive experiments demonstrating that TransMamba achieves superior training efficiency and performance compared to baselines, and validated the deeper consistency between Transformer and Mamba paradigms, offering a scalable solution for next-generation sequence modeling.

Xingwu Sun, Shuaipeng Li, Ruobing Xie et al.

Low-precision training is considered an effective strategy for reducing both training and downstream inference costs. Previous scaling laws for precision mainly focus on integer quantization, which pay less attention to the constituents in floating-point (FP) quantization, and thus cannot well fit the LLM losses in this scenario. In contrast, while FP quantization training is more commonly implemented in production, it's research has been relatively superficial. In this paper, we thoroughly explore the effects of FP quantization targets, exponent bits, mantissa bits, and the calculation granularity of the scaling factor in FP quantization training performance of LLM models. In addition to an accurate FP quantization unified scaling law, we also provide valuable suggestions for the community: (1) Exponent bits contribute slightly more to the model performance than mantissa bits. We provide the optimal exponent-mantissa bit ratio for different bit numbers, which is available for future reference by hardware manufacturers; (2) We discover the formation of the critical data size in low-precision LLM training. Too much training data exceeding the critical data size will inversely bring in degradation of LLM performance; (3) The optimal FP quantization precision is directly proportional to the computational power, but within a wide computational power range. We estimate that the best cost-performance precision should lie between 4-8 bits.

Weiyan Wang, Xingjian Shi, Ruiqi Shu et al.

In practice, physical spatiotemporal forecasting can suffer from data scarcity, because collecting large-scale data is non-trivial, especially for extreme events. Hence, we propose \method{}, a novel probabilistic framework to realize iterative self-training with new self-ensemble strategies, achieving better physical consistency and generalization on extreme events. Following any base forecasting model, we can encode its deterministic outputs into a latent space and retrieve multiple codebook entries to generate probabilistic outputs. Then BeamVQ extends the beam search from discrete spaces to the continuous state spaces in this field. We can further employ domain-specific metrics (e.g., Critical Success Index for extreme events) to filter out the top-k candidates and develop the new self-ensemble strategy by combining the high-quality candidates. The self-ensemble can not only improve the inference quality and robustness but also iteratively augment the training datasets during continuous self-training. Consequently, BeamVQ realizes the exploration of rare but critical phenomena beyond the original dataset. Comprehensive experiments on different benchmarks and backbones show that BeamVQ consistently reduces forecasting MSE (up to 39%), enhancing extreme events detection and proving its effectiveness in handling data scarcity.

Ang Lv, Ruobing Xie, Shuaipeng Li et al.

We propose a novel attention mechanism, named Cog Attention, that enables attention weights to be negative for enhanced expressiveness, which stems from two key factors: (1) Cog Attention enhances parameter flexibility. For example, unlike traditional softmax attention heads that use a static output-value (OV) matrix to delete or copy inputs that the heads attend to, Cog Attention naturally learns to use the sign of dynamic query-key (QK) inner products to represent these operations. This enables Cog Attention to perform multiple operations simultaneously within a single head. Meanwhile, Cog Attention's OV matrix can focus more on refinement or modification. (2) Cog Attention enhances the model's robustness against representational collapse by preventing the ``over-squashing'' of earlier tokens into later positions. We develop Transformer-like models which use Cog Attention as attention modules, including decoder-only models at various scales for language modeling and U-ViT diffusion models for image generation. Experiments show that models using Cog Attention exhibit superior performance compared to those employing traditional softmax attention modules. Our approach suggests a promising research direction for rethinking and breaking the entrenched constraints of traditional softmax attention, such as the requirement for non-negative weights.

Hao Wu, Yuan Gao, Xingjian Shi et al.

To address the dual challenges of inherent stochasticity and non-differentiable metrics in physical spatiotemporal forecasting, we propose Spatiotemporal Forecasting as Planning (SFP), a new paradigm grounded in Model-Based Reinforcement Learning. SFP constructs a novel Generative World Model to simulate diverse, high-fidelity future states, enabling an "imagination-based" environmental simulation. Within this framework, a base forecasting model acts as an agent, guided by a beam search-based planning algorithm that leverages non-differentiable domain metrics as reward signals to explore high-return future sequences. These identified high-reward candidates then serve as pseudo-labels to continuously optimize the agent's policy through iterative self-training, significantly reducing prediction error and demonstrating exceptional performance on critical domain metrics like capturing extreme events.

Guoliang Zhao, Yuhan Fu, Shuaipeng Li et al. · tsinghua

Mixture-of-Experts (MoE) models have become the consensus approach for enabling parameter-efficient scaling and cost-effective deployment in large language models. However, existing scaling laws for dense models are inapplicable to MoE models, which stems from three critical challenges: the multiplicity of influencing factors, their intricate coupling relationships and the non-monotonic nature of their performance impacts. They collectively necessitate a fine-grained investigation into MoE-specific scaling laws. In this work, we perform a systematic decomposition of MoE settings, identifying five key factors that influence model performance from both size and structural perspectives (data size ($D$), total model size ($N$), activated model size ($N_a$), number of active experts ($G$) and the ratio of shared experts ($S$)). Specifically, we design $446$ controlled experiments to characterize their marginal effects, ultimately constructing a comprehensive and precise joint MoE scaling law that considers all essential factors. Furthermore, we derive the theoretically optimal and practically efficiency-aware optimal configurations for $G$, $S$ and $N_a/N$ with detailed analyses. Our results demonstrate that the optimal settings for $G$ and $S$ are independent of both the model architecture and data size. With the scaling of $N$, the optimal activation parameter ratio of $N_a/N$ becomes sparser. Our proposed MoE scaling law could function as an accurate and insightful guidance to facilitate future MoE model design and training.

Hao Wu, Yuan Gao, Ruijian Gou et al.

Reliable long-term forecasting of Earth system dynamics is fundamentally limited by instabilities in current artificial intelligence (AI) models during extended autoregressive simulations. These failures often originate from inherent spectral bias, leading to inadequate representation of critical high-frequency, small-scale processes and subsequent uncontrolled error amplification. Inspired by the nested grids in numerical models used to resolve small scales, we present TritonCast. At the core of its design is a dedicated latent dynamical core, which ensures the long-term stability of the macro-evolution at a coarse scale. An outer structure then fuses this stable trend with fine-grained local details. This design effectively mitigates the spectral bias caused by cross-scale interactions. In atmospheric science, it achieves state-of-the-art accuracy on the WeatherBench 2 benchmark while demonstrating exceptional long-term stability: executing year-long autoregressive global forecasts and completing multi-year climate simulations that span the entire available $2500$-day test period without drift. In oceanography, it extends skillful eddy forecast to $120$ days and exhibits unprecedented zero-shot cross-resolution generalization. Ablation studies reveal that this performance stems from the synergistic interplay of the architecture's core components. TritonCast thus offers a promising pathway towards a new generation of trustworthy, AI-driven simulations. This significant advance has the potential to accelerate discovery in climate and Earth system science, enabling more reliable long-term forecasting and deeper insights into complex geophysical dynamics.

Fangyu Liu, Rongtian Ye, Xun Wang et al.

The hubness problem widely exists in high-dimensional embedding space and is a fundamental source of error for cross-modal matching tasks. In this work, we study the emergence of hubs in Visual Semantic Embeddings (VSE) with application to text-image matching. We analyze the pros and cons of two widely adopted optimization objectives for training VSE and propose a novel hubness-aware loss function (HAL) that addresses previous methods' defects. Unlike (Faghri et al.2018) which simply takes the hardest sample within a mini-batch, HAL takes all samples into account, using both local and global statistics to scale up the weights of "hubs". We experiment our method with various configurations of model architectures and datasets. The method exhibits exceptionally good robustness and brings consistent improvement on the task of text-image matching across all settings. Specifically, under the same model architectures as (Faghri et al. 2018) and (Lee at al. 2018), by switching only the learning objective, we report a maximum R@1improvement of 7.4% on MS-COCO and 8.3% on Flickr30k.

Fangyu Liu, Shuaipeng Li, Liqiang Zhang et al.

Semantic parsing of large-scale 3D point clouds is an important research topic in computer vision and remote sensing fields. Most existing approaches utilize hand-crafted features for each modality independently and combine them in a heuristic manner. They often fail to consider the consistency and complementary information among features adequately, which makes them difficult to capture high-level semantic structures. The features learned by most of the current deep learning methods can obtain high-quality image classification results. However, these methods are hard to be applied to recognize 3D point clouds due to unorganized distribution and various point density of data. In this paper, we propose a 3DCNN-DQN-RNN method which fuses the 3D convolutional neural network (CNN), Deep Q-Network (DQN) and Residual recurrent neural network (RNN) for an efficient semantic parsing of large-scale 3D point clouds. In our method, an eye window under control of the 3D CNN and DQN can localize and segment the points of the object class efficiently. The 3D CNN and Residual RNN further extract robust and discriminative features of the points in the eye window, and thus greatly enhance the parsing accuracy of large-scale point clouds. Our method provides an automatic process that maps the raw data to the classification results. It also integrates object localization, segmentation and classification into one framework. Experimental results demonstrate that the proposed method outperforms the state-of-the-art point cloud classification methods.