Search AI Products and News

Meta has released its latest open-source AI model, Llama 4, marking another significant advancement in the field of artificial intelligence. Llama 4 comes in two versions, Scout and Maverick, designed to enhance AI model capabilities and performance. Meta states that Llama 4 is a multimodal large language model capable of processing various data types, including text, images, video, and audio, and can freely convert between these formats. Notably, the Llama 4 series is the first...

Ant Group's Ling team recently published a preprint on arXiv titled "Every FLOP Matters: Scaling a 300-billion parameter Mixture-of-Experts LING model without high-end GPUs," detailing two novel large language models: Ling-Lite and Ling-Plus. These models incorporate several innovations enabling efficient training on low-performance hardware, significantly reducing training costs.

Tencent recently released the official version of its HunYuan large language model series – HunYuan-T1. This new model, built upon the medium-scale HunYuan base model and extensively fine-tuned, demonstrates significantly improved reasoning capabilities, particularly excelling in deep thinking and complex problem-solving. Since the launch of the HunYuan T1-Preview in February, users have experienced faster and more insightful processing. The official release marks a significant upgrade to the product line. The HunYuan-T1 development team leveraged the latest Turbo...

ByteDance's Doubao large model team recently announced a breakthrough in addressing key bottlenecks in Mixture-of-Experts (MoE) architecture, open-sourcing a significant optimization technology called COMET. This technology dramatically improves large model training efficiency, achieving a remarkable 1.7x speedup and a 40% reduction in training costs. Image Note: Image generated by AI, image licensing provider Midjourney. COMET has been deployed in ByteDance's multi-thousand-GPU cluster training, resulting in millions of GPU hours saved.

DeepSeek announced its second-day product of the open-source week: the first open-source EP communication library for MoE models, enabling full-stack optimization for mixed-expert model training and inference. DeepEP is a high-efficiency communication library specifically designed for Mixture-of-Experts (MoE) and Expert Parallelism (EP). It aims to provide high-throughput and low-latency many-to-many GPU kernel communication, commonly known as MoE routing and aggregation. DeepEP not only supports low-precision operations such as FP8 but also integrates with the DeepSeek-V3 paper.

The ByteDance Doubao large model team announced today the successful development of a new sparse model architecture called UltraMem. This architecture effectively addresses the high memory access issues during the inference of MoE (Mixture of Experts) models, improving inference speed by 2 to 6 times compared to MoE, and reducing inference costs by up to 83%. This groundbreaking advancement opens a new path for efficient inference of large models. The UltraMem architecture successfully resolves the memory bottleneck during inference of MoE architectures while maintaining model performance. Experimental results show that the parameters and activation conditions are the same.

ByteDance officially launches its latest Doubao large model 1.5 Pro (Doubao-1.5-pro), which demonstrates outstanding comprehensive capabilities in various fields, successfully surpassing the well-known GPT-4o and Claude3.5Sonnet in the industry. The release of this model marks an important step forward for ByteDance in the field of artificial intelligence. Doubao 1.5 Pro adopts a novel sparse MoE (Mixture of Experts) architecture, utilizing a smaller set of activation parameters for pre-training. This design's innovation...

As a new wave of competition arises in the global AI video generation field, Chinese tech company MiniMax's 'Haibao AI' has made a strong impression with its remarkable performance. Data shows that this AI product, which was launched in April of this year, achieved a staggering 867.41% increase in visit volume in September, rapidly becoming the top product in global AI visit growth rankings. The core competitiveness of this product lies in its underlying technological support. MiniMax built Haibao AI based on the trillion-parameter MoE model abab-6.5, allowing it to compete with GPT-4.

Today, OPPO announced that it has successfully applied the Expert Mixture of Experts (MoE) architecture to mobile devices for the first time globally. This groundbreaking advancement significantly enhances AI processing efficiency, opening new pathways for the further development of terminal AI technologies and laying the foundation for the deep integration of AI and mobile hardware. As AI technology rapidly advances, an increasing number of AI tasks need to be completed on terminal devices. However, large AI models often require powerful computing capabilities, which poses a huge challenge for mobile devices with limited hardware resources. This breakthrough aims to overcome this limitation.

Shenzhen Yuanxiang Information Technology Co., Ltd. recently announced the successful release of China’s largest Mixture of Experts (MoE) open source large model — XVERSE-MoE-A36B. The launch of this model marks a significant advancement in China’s AI field, elevating domestic open source technology to an internationally leading level.

The Allen Institute for Artificial Intelligence (AI2) has released an open-source large language model OLMoE, aimed at providing high performance and low-cost solutions. This model employs a Sparse Mixture of Experts (MoE) architecture with 7 billion parameters, but through an intelligent routing mechanism, only 1 billion parameters are used per input token, achieving efficient computation. OLMoE includes both general and instruction-tuned versions, supporting a context window of 4096 tokens. Its training data is sourced from a wide range, including Common Crawl, Dolma CC, and Wikipedia.

At the 2024 Tencent Global Digital Ecosystem Conference, Tencent's Vice President and Head of Cloud Business, Qiu Yuepeng, announced the official launch of the Tencent Hunyuan Turbo large model. This new generation of large models is designed based on the MoE (Mixture of Experts) architecture, achieving a doubling in inference efficiency with a 100% improvement compared to the previous generation, while significantly reducing inference costs by 50%. Additionally, Hunyuan Turbo also demonstrates outstanding performance in decoding efficiency, improving by 2.

MiniMax Xiyu Technology recently announced that since its establishment at the end of 2021, it has built trillion-parameter MoE large text models, speech models, and image models. It has also launched native applications such as Xingye and Hailuo AI, while providing an open platform API service for enterprises and developers.

MoE Technology has announced the open-source release of its independently developed audio understanding large model, MooER, which is the first large-scale open-source speech model trained and inferred on domestically produced full-featured GPUs. MooER was trained on a large-scale audio dataset in just 38 hours on the MoE KuaE Intelligence Computing Platform, demonstrating excellent performance in Chinese and English speech recognition, as well as Chinese-to-English speech translation, achieving an impressive BLEU score of 25.2 in the Covost2 Chinese-to-English test set, nearing industrial-grade effectiveness. MoE Technology plans to further open-source the training code.

MoeThread's AI creation platform 'MoBi MaLiang' announces a complete upgrade, providing users with a more efficient and personalized image generation experience. The upgraded MoBi MaLiang has particularly enhanced its understanding of Chinese language and culture, making creative portrait generation more precise. Users can generate personalized creative portraits quickly in just three simple steps: upload a photo, enter a description, and generate an image.

On January 16, 2024, MiniMax released China's first MoE large language model, abab6. The MoE architecture enables abab6 to handle complex tasks and train more data in a shorter time. abab6 outperforms the previous version abab5.5 in instruction adherence, Chinese comprehension, and English comprehension. abab6 demonstrates excellent capabilities, such as teaching children math problems and helping to build a fictional board game about Shanghai.