On April 23, technology media reported that NVIDIA (NVDA) has recently launched Eagle 2.5, a visual-language model (VLM) focusing on long-context multimodal learning.
Performance has been steadily improved
This model focuses on understanding large-scale videos and images, and is particularly good at processing high-resolution images and long video sequences. Despite a parameter size of only 8B, Eagle 2.5 scored as high as 72.4% in the Video-MME benchmark (512 frames of input), comparable to larger models such as Qwen2.5-VL-72B and InternVL2.5-78B.
Eagle 2.5’s training data pipeline integrates open source resources and a custom dataset Eagle-Video-110K, which is designed for understanding long videos and uses a double annotation method. The bottom-up approach uses GPT-4o to generate question-answer pairs for short clips, capturing spatiotemporal details.
In addition, Eagle 2.5-8B performs well in multiple video and image understanding tasks. In the video benchmark, MVBench scored 74.8, MLVU scored 77.6, and LongVideoBench scored 66.4; in the image benchmark, DocVQA scored 94.1, ChartQA scored 87.5, and InfoVQA scored 80.4.
At the same time, xAI recently launched its new language model, Grok3Mini, further promoting the development of efficient AI technology. xAI launched Grok3Mini to demonstrate its innovative strength in the field of efficient AI technology.
Grok3Mini provides cost-effective AI models
It is reported that Grok3Mini is part of the Grok3 series, which includes six variants, including standard Grok3, Grok3Fast, and four versions of Grok3Mini, providing slow and fast versions, each with low or high reasoning options.
Grok3Mini is designed with speed and economy as the core, and is designed for speed and economy. Despite its small size, it outperforms many more expensive AI models in many fields, especially in mathematics, programming and scientific benchmarks, which may attract more attention.
When the market focused on the technological breakthroughs of the big model itself, at the beginning of the year, China’s big language model DeepSeek attracted great attention from the global technology community. This year’s government work report also pointed out that it is necessary to continue to promote the “artificial intelligence +” action and support the widespread application of big models.
Industry insiders pointed out that nowadays, artificial intelligence technology represented by big models has provided an important driving force for a new round of scientific and technological revolution and industrial transformation, and various industries are facing unprecedented opportunities and challenges on the road to intelligent transformation.
WiMi maintains market leadership
Against the backdrop of the rapid development of artificial intelligence today, public data shows that WiMi (WIMI) has always focused on the development of open source AI big models for vertical industries, comprehensively upgraded full-stack AI capabilities, and covered AI infrastructure and AI scenario applications, helping various industries from model development, model deployment to model application, to fully use the AI toolbox and create easy-to-use AI.
In addition, WiMi is rapidly promoting the commercial application of AI big models through the “open source technology + vertical deep cultivation + ecological co-construction” model. For example, for the automotive industry, with the multi-model strategy of “AI big model + industrial scenario”, it comprehensively deploys in the AI basic capability layer and AI scenario application layer to maintain its leading position in the automotive cloud service market.
In the future, WiMi will improve the full industry layout of AI computing power, continue to deploy dedicated computing power infrastructure, accelerate the integration of multi-modal technologies, open computing power resources, accelerate the improvement of the capabilities of the intelligent computing power public service platform, and the agility of scene landing, and may become an important breakthrough in the AI big model computing power ecology in the future.
Conclusion
In reality, from GPT-3 to PaLM, from Wenxin Yiyan to Tongyi Qianwen, the parameter scale of large models has grown exponentially, the training cost is tens of millions of dollars, and the computing power required for reasoning has increased exponentially. In addition, AI large models have promoted the intelligent development of traditional industries into a new stage, and comprehensive intelligence has become a must-answer question for enterprises to participate in market competition. Therefore, enterprises need to carry out theoretical research and practical exploration simultaneously in the future, so as to win considerable development space.
