Home Innovation Cloud Alibaba Cloud Boosts AI with N...

Alibaba Cloud Boosts AI with New Models, Infrastructure

Cloud

Business Honor
17 April, 2025

With additional LLMs, GenAI tools, and improved infrastructure support, Alibaba Cloud expands its AI solutions.

Alibaba Cloud has strengthened its artificial intelligence (AI) capabilities for clients globally by introducing a variety of new models, tools, and infrastructure improvements.

The new capabilities, which were unveiled at the company's Spring Launch 2025 online event, are centered on offering scalable AI solutions, especially in the quickly developing fields of large language models (LLMs) and generative AI (GenAI).

Expanded access to Alibaba Cloud's core models, including the most recent versions of its proprietary LLM series, Qwen, is a key component of the announcement. The modal model Qwen2.5-Omni-7b, the logic model QwQ-Plus, the visual reasoning model QVQ-Max, and the massive amounts mixture of experts (MoE) model Qwen-Max are among the models that are available through the company's availability zones in Singapore.

With algorithm-driven accuracy, QwQ-Plus tackles intricate question-and-answer assignments and expert-level mathematics challenges, fostering deep analytical thinking. In contrast, QVQ-Max concentrates on visual reasoning and can solve intricate multimodal problems with greater precision and reasoning power. It supports visual inputs as well as reasoning-based outputs.

Platform for AI (PAI), Alibaba Cloud's machine learning platform, has been significantly improved to accommodate these complex models. With its multi-node design and distributed inference capabilities, PAI-Elastic Algorithm Service (EAS) can handle the growing demands of super-large models, especially those that use MoE structures and ultra-long-text processing.

Additionally, PAI-EAS lowers costs and improves performance by introducing a prefill-decode disaggregation mechanism. According to Alibaba Cloud, when implemented using the Qwen2.5-72B model, this innovation boosts concurrency by 92% and tokens per second by 91%.

Another significant update was made to PAI-Model Gallery, which currently offers close to 300 open-source models. This covers the entire line of Alibaba Cloud's open-source Qwen and Wan series, which can all be accessed via a deployment and management interface that doesn't require any code.

Along with new capabilities like model evaluation for performance insights and model distillation to lower deployment costs, the gallery provides a variety of deployment techniques and underlying processing resources.