Search AI Products and News

Recently, the AI code generation field has seen a surge in open-source releases, with several heavyweight models making their debut. Among them, the Cogito v1 Preview series from Deep Cogito stands out. According to AIbase, this new family of open-source models includes various sizes: 3B, 8B, 14B, 32B, and 70B parameters. Not only does it outperform competitors in its class, but its 70B version even surpasses Meta's recently released Llama 4 109B MoE model, sparking considerable industry discussion.

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.

A collaborative team from Ant Group, Zhejiang University, University of Liverpool, and East China Normal University presented ScaleOT, a novel cross-domain fine-tuning framework, at AAAI 2025. This framework achieves a 50% improvement in privacy protection and a remarkable 90% reduction in computational cost compared to knowledge distillation, without compromising model performance. ScaleOT offers a highly efficient and lightweight solution for cross-domain fine-tuning of billion-parameter models. The innovative nature of the paper secured it an oral presentation at AAAI.

With the rapid development of artificial intelligence, the integration of visual and language capabilities has led to groundbreaking advancements in visual language models (VLMs). These models aim to simultaneously process and understand both visual and textual data, being widely applied in scenarios such as image captioning, visual question answering, optical character recognition, and multimodal content analysis. VLMs play a significant role in developing autonomous systems, enhancing human-computer interaction, and creating efficient document processing tools, successfully bridging the gap between these two data modalities. However, there are challenges in handling high-resolution visual data and diverse textual input.

Recently, research teams from several Chinese scientific institutions have launched the super large-scale multimodal dataset named Infinity-MM, and based on this dataset, they have trained an outstanding AI model called Aquila-VL-2B. This breakthrough injects new momentum into the development of multimodal AI. The Infinity-MM dataset is astonishing in size, containing four major types of data: 10 million image descriptions, 24.4 million general visual instruction data, 6 million selected high-quality instruction data, and 3 million generated by G

Meta CEO Mark Zuckerberg recently announced the official release of Llama3.1, stating that it could be a significant milestone in the open-source AI field. Meta has introduced three different sizes of Llama3.1 models, with the most notable being the large model with 405 billion parameters. Zuckerberg claims it to be the "most advanced open-source model" currently available, and it can even compete with leading closed-source models in certain areas.In addition to the 405 billion parameter model,

小冰公司的最新发布标志着其AI数字员工产品的重大升级，新增“零样本”数字人技术（Zero-shot Xiaoice Neural Rendering，Zero-XNR）、超千亿大模型基座以及高能多媒体传输系统，显著提升实时交互质量和效果。Zero-XNR技术结合TTS语音模型，并配以高效聚类框架实现秒级高质量声音与形象复刻。大模型基座与Agent构建框架的强化集结了强大的职业交互功能，用于精准商业交互。透影音画传输系统的引入，则确保了超高清视频的传输流畅性与抗干扰性，增强了用户体验。在技术创新引领下，小冰旨在推动数字人技术的普及与产业化应用，经过微软背景的专业打磨，现已成为一家独享核心技术的中国AI公司。其技术框架涵盖了自然语言处理、语音识别、视觉交互及AI内容生成，已成功构建全球化的AI数字人产品体系，服务从金融到教育、智能汽车、智能地产等众多行业，支持数字工作者提供平稳、可靠并生产力充足的交互体验。

In this era of information overload, we interact with smart devices every day. Have you ever wondered how these seemingly intelligent creatures know to "take an umbrella because it's raining"? Behind this is a profound transformation in causal reasoning.A group of researchers from renowned academic institutions such as Microsoft and MIT have jointly developed a groundbreaking machine learning training strategy. This strategy not only overcomes the shortcomings of large machine learning models in

Nvidia recently announced the launch of two groundbreaking AI technologies: HelpSteer2 and Llama3-70B-SteerLM-RM. These innovations promise to significantly enhance the capabilities of AI systems across a variety of applications, from autonomous driving to natural language processing.

Recently, Moore Threads and the all-subject educational AI model 'Shizhe AI' jointly announced that they have completed the training and testing of their large model. Utilizing Moore Threads' KUAE supercomputing cluster, Shizhe AI successfully completed the intensive training of a 7 billion parameter model in one week, achieving the expected training efficiency and fully demonstrating the capabilities of the domestically produced all-functional GPU for trillion-scale training.

In the early hours of today, Alibaba's Tongyi Qianwen team released the Qwen2 series of open-source models. This series includes pre-trained and instruction-tuned models in five sizes: Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, Qwen2-57B-A14B, and Qwen2-72B. Key information indicates that these models show significant improvements in parameter count and performance compared to the previous generation Qwen1.5.

Scientists successfully trained a model at ChatGPT level using only 8% of the computing power. The research team utilized innovative technology on the Frontier supercomputer, training a trillion-parameter language model with just thousands of AMD GPUs. They achieved 100% weak scaling efficiency for 175 billion-parameter and trillion-parameter models while utilizing only 8% of computing capacity. The results indicate that training large language models with trillion parameters remains challenging.

Nucleus AI has launched a 22 billion parameter language model under the open-source MIT license, positioned between the 13B and 34B models. Nucleus AI plans to build an OS centered on agriculture, using AI research to super-optimize supply and demand. The CEO of Nucleus AI stated that they plan to release various sizes of language models as well as two RetNet models. Nucleus AI aims to leverage AI to optimize agricultural supply and demand to address various challenges faced by farmers.

California-based startup Nucleus AI has released a large-scale language model with 2.2 billion parameters, aimed at AI transformation in the agricultural sector. This model is available under open-source MIT licensing and commercial licensing, with superior performance suitable for various generation tasks. Nucleus plans to release more versions of the 22B model, as well as two RetNet models, which will enhance speed and reduce costs. Nucleus AI's vision is to build an intelligent agricultural operating system to optimize supply and demand relationships and address multiple challenges in the agricultural sector.

SambaNova has launched its intelligent AI chip SN40L, which can run models with up to 50 trillion parameters while maintaining model accuracy. The SN40L chip is manufactured by TSMC and provides over 256k sequence length for a single system node, enhancing model quality and inference speed through integrated technology. SambaNova's full-stack large language model (LLM) platform will be powered by the SN40L chip, addressing the challenges enterprises face in deploying generative artificial intelligence. The uniqueness of the SN40L chip lies in its simultaneous processing capabilities.

Researchers from Tsinghua University and others have introduced MathGLM, a 2 billion parameter language model that outperforms GPT-4 with nearly 100% accuracy in arithmetic tasks. MathGLM utilizes a decoder-only architecture and has significantly improved its mathematical capabilities through training on a large-scale arithmetic dataset. It also surpasses models like GPT-4 and ChatGPT in handling complex mixed arithmetic operations with intricate numerical formats.

1. A 100 billion parameter model requires 266 8-card A100 servers with a single card computing efficiency of 44%. 2. To improve the performance of large models, it is necessary to optimize aspects such as frameworks, IO, and communication. 3. Compared to GPT-4, domestic large models have discrepancies in computing power, algorithms, and data.