Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought

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.

A Hierarchical Location Normalization System for Text

It's natural these days for people to know the local events from massive documents. Many texts contain location information, such as city name or road name, which is always incomplete or latent. It's significant to extract the administrative area of the text and organize the hierarchy of area, called location normalization. Existing detecting location systems either exclude hierarchical normalization or present only a few specific regions. We propose a system named ROIBase that normalizes the text by the Chinese hierarchical administrative divisions. ROIBase adopts a co-occurrence constraint as the basic framework to score the hit of the administrative area, achieves the inference by special embeddings, and expands the recall by the ROI (region of interest). It has high efficiency and interpretability because it mainly establishes on the definite knowledge and has less complex logic than the supervised models. We demonstrate that ROIBase achieves better performance against feasible solutions and is useful as a strong support system for location normalization.

A Dynamic Evolutionary Framework for Timeline Generation based on Distributed Representations

Given the collection of timestamped web documents related to the evolving topic, timeline summarization (TS) highlights its most important events in the form of relevant summaries to represent the development of a topic over time. Most of the previous work focuses on fully-observable ranking models and depends on hand-designed features or complex mechanisms that may not generalize well. We present a novel dynamic framework for evolutionary timeline generation leveraging distributed representations, which dynamically finds the most likely sequence of evolutionary summaries in the timeline, called the Viterbi timeline, and reduces the impact of events that irrelevant or repeated to the topic. The assumptions of the coherence and the global view run through our model. We explore adjacent relevance to constrain timeline coherence and make sure the events evolve on the same topic with a global view. Experimental results demonstrate that our framework is feasible to extract summaries for timeline generation, outperforms various competitive baselines, and achieves the state-of-the-art performance as an unsupervised approach.