Alibaba has announced a set of upgrades across its AI products, spanning cloud infrastructure, model services, chips and foundation models, as the company targets growing demand for “agentic” AI systems that can carry out multi-step tasks.
The announcements were made at the Alibaba Cloud Summit and include Qwen3.7-Max, described by the company as its latest large language model aimed at agentic coding, complex reasoning and long-horizon task execution. Alibaba said Qwen3.7-Max will be available soon for developers and enterprises worldwide.
Alongside the model, Alibaba Cloud introduced infrastructure updates intended to support higher AI workload requirements. These include the Panjiu AL128 Supernode Server, which the company said is designed for scalable agent inference and large-scale model training, and an optimisation update within its model service platform intended to refine model performance over time.
Alibaba’s semiconductor design subsidiary T-Head also announced the Zhenwu M890 AI training and inference processor, which the company said includes 144GB of GPU memory, 800GB per second of inter-chip bandwidth, and support for multiple precision formats down to FP4. T-Head also unveiled ICN Switch 1.0, a dedicated switching chip it said delivers up to 25.6 Tbps of aggregate bandwidth.
The company said the Panjiu AL128, powered by the Zhenwu M890 and ICN Switch 1.0, integrates 128 AI accelerators within a single rack. Alibaba said the system is now available on its model service platform Model Studio for the China market, referred to as “Bailian”.
For the security and governance implications of increasingly autonomous systems, Alibaba said Bailian includes “built-in safety governance capabilities” intended to keep autonomously operating agents within defined boundaries. It also announced “Agentic RL”, a reinforcement learning mechanism that it said uses agent execution feedback to support ongoing model iteration.
Alibaba also made performance claims about Qwen3.7-Max’s ability to run extended tasks, stating the model can “sustain continuous operation for up to 35 hours” and manage more than 1,000 tool calls. It said the model is optimised for agent frameworks including OpenClaw, Hermes Agent, Claude Code, Qwen Paw and Qoder, and will be accessible through Model Studio for global developers.
T-Head said it has delivered more than 560,000 Zhenwu units to date and that external customers across 20 industries have deployed the chips.
