Alibaba Cloud announced new large models and infrastructure upgrades at its annual Apsara Conference 2025.
The updates cover language and multimodal models, agent development frameworks, enterprise AI platforms, and cloud infrastructure.
“In the future, large AI models will be deeply integrated into a wide range of devices, functioning like operating systems, equipped with persistent memory, seamless cloud-edge coordination, and the ability to continuously evolve,”
said Eddie Wu, Chairman and CEO of Alibaba Cloud Intelligence.
He added that Alibaba Cloud remains committed to open-sourcing Qwen and shaping it into the “operating system of the AI era,” while continuing with a RMB 380 billion investment in AI and cloud infrastructure over the next three years.
Since launching Qwen in 2023, Alibaba has open-sourced more than 300 models built on its Qwen language and Wan visual-generation foundations, with over 600 million downloads and 170,000 derivative models created.
More than one million users have accessed Qwen through Model Studio, its AI development platform.
The company released Qwen3-Max, its largest LLM to date with over 1 trillion parameters.
Available in Instruct and Thinking modes, it scored 69.6 on the SWE-Bench benchmark for software issue resolution and showed strong performance on Tau2-Bench, which tests conversational agent capabilities.
Other models include Qwen3-VL, a vision-language model with spatial understanding for 3D navigation and visual programming; Qwen3-Omni, a multimodal model that processes text, images, audio, and video with real-time responses and is now open-sourced on Hugging Face and ModelScope; and upgraded versions of Qwen3-Coder and Qwen3-Image-Edit.
Alibaba also introduced Fun-ASR, a speech recognition model, and Fun-CosyVoice, a speech synthesis model.
The company also previewed Wan2.5 models for video, image generation, and editing.
The video models now generate clips of up to 10 seconds with integrated audio and improved multimodal alignment.
Model Studio has been updated with a new Agent Development Kit (ADK) for high-code enterprise projects and an enhanced low-code Agent Development Platform (ADP).
More than 800,000 agents have been created on Model Studio.
Other platforms introduced include AgentBay, with self-evolving functions and compliance controls; Lingyang AgentOne, an enterprise AI application platform that integrates marketing, analytics, and operations; and Zaodian, a consumer-facing service for AI-generated images and videos.
Alibaba Cloud also introduced infrastructure upgrades to support large-scale AI and agentic workloads.
Its Object Storage Service now includes vector storage for retrieval-augmented generation applications.
The new HPN8.0 network architecture delivers 800 Gbps throughput for large model training and inference.
Its Cloud Threat Detection Response system has been enhanced with agent-driven automation, increasing automated investigation success from 59% to 74%.
Container Compute Services now support scaling up to 15,000 pods per minute with stronger isolation features.
PolarDB has been upgraded with Compute Express Link (CXL) to reduce latency and support multimodal data storage through a hybrid Lakebase architecture.
AThe Platform for AI now offers optimisations that accelerate Qwen training by over 300% and reduce Wan training time by 28%, while also improving inference efficiency.
Featured image credit: Alibaba Cloud