Alibaba debuts Qwen 3.7‑Max and Zhenwu M890 to power agentic AI workloads

At its Hangzhou cloud summit, Alibaba introduced the Qwen 3.7‑Max agent‑focused LLM, the Zhenwu M890 AI processor and the Panjiu AL128 rack server, plus platform updates aimed at sustained, multi‑step agent tasks and commercial deployment.

Alibaba used its cloud summit in Hangzhou to unveil a package of hardware, software and model updates aimed at agentic AI: the Qwen 3.7‑Max large language model, the T‑Head Zhenwu M890 processor, the Panjiu AL128 Supernode Server, and platform enhancements on its Bailian (Model Studio) service. The company demonstrated the stack with an extended unattended engineering test in which Qwen 3. manufacturer’s official version by tenfold — an endurance and integration signal for builders and operators of continuous agent workflows.

Qwen 3.7‑Max is positioned for sustained, multi‑step operations rather than single‑turn conversational replies. Alibaba says the model is engineered to manage multi‑file software projects, orchestrate multi‑agent office workflows and integrate with agent harnesses such as OpenClaw, Hermes Agent, Claude Code, Qwen Paw and Qoder. The emphasis is on long‑running operational continuity and heavy tool use, addressing common failure modes in chained, agentic tasks.

To host concurrent, bursty agent workloads, Alibaba introduced the Panjiu AL128 Supernode Server, a rack‑scale system that integrates 128 AI accelerators and delivers petabyte‑per‑second internal bandwidth. The server is available through the company’s Bailian model service platform (rebranded as Model Studio outside China) and is paired with platform features intended to close the loop between live agent behavior and model refinement — most notably an Agentic RL reinforcement learning mechanism and built‑in safety governance guardrails.

T‑Head detailed the Zhenwu M890 processor that it says yields three times the performance of the prior Zhenwu 810E. The M890 includes 144 GB of on‑chip memory, 800 GB/s inter‑chip bandwidth and native support for numerical precisions ranging from FP32 down to FP4. It is paired with an ICN Switch 1.0 networking chip that provides up to 25.6 Tbps of aggregate bandwidth, and is supported by the T‑Head SAIL™ software stack designed to maximize hardware utilization. Alibaba presented these components as the compute and network foundation for the Panjiu AL128 Supernode.

Alibaba also framed the launch with operational and commercial metrics: T‑Head has shipped more than 560,000 Zhenwu chips to date, and the company says over 400 external customers across 20 industries — including automakers and financial services firms — have deployed them. Executives projected that the model and application services platform will exceed RMB 10 billion (about US$1.4 billion) in annual recurring revenue in the June quarter and reach RMB 30 billion (about US$4.1 billion) by year‑end, signaling an effort to monetize agentic AI at scale.

For developers and infrastructure teams, the combined releases target specific bottlenecks: longer‑running agent models require large on‑chip memory, high interconnect bandwidth and congestion‑free cluster networking, which Panjiu AL128 and the ICN Switch 1.0 aim to provide. Bailian’s Agentic RL and safety guardrails are meant to let teams iterate models from real‑world agent outcomes while retaining operational controls, and the M890’s FP4 — FP32 support offers flexibility between inference cost and training accuracy.