Alibaba is placing a bigger bet on AI that does the work, not just answers the question. At its Cloud Summit in Hangzhou, the company laid out a top-to-bottom upgrade of its AI stack built for what it calls the agentic era, led by a new flagship model, Qwen3.7-Max, and a homegrown AI chip. For Malaysian developers and businesses that build on Alibaba Cloud, it signals where the platform they rent is heading.
Editor
Kai T chevron_right
Table of Contents
What Alibaba's Qwen3.7-Max brings
Qwen3.7-Max is Alibaba's latest large language model, tuned for agentic coding, complex reasoning, and long-horizon tasks, the kind of multi-step jobs an AI agent runs on its own rather than in a single back-and-forth. Its context window has grown to 1 million tokens, up from 256,000 on the previous Qwen3.6-Max-Preview, which gives it far more room to hold a large codebase or a long task in working memory.
Alibaba illustrated the point with an internal benchmark: it says Qwen3.7-Max ran for 35 hours straight, made more than 1,000 tool calls, and produced a production-grade computing kernel that beat the chip maker's official version by tenfold. Those are Alibaba's own figures and have not been independently verified, but the demonstration is the message: a model meant to keep working without a human in the loop.
The chips under the model
The upgrade also reaches down to silicon designed by T-Head, Alibaba's chip unit. The new Zhenwu M890 accelerator delivers three times the performance of its predecessor, the Zhenwu 810E, with 144GB of on-chip memory, 800GB per second of inter-chip bandwidth, and support for number formats from FP32 down to FP4. It is paired with the ICN Switch 1.0, which Alibaba says moves 25.6 Tbps across clusters of 64 accelerators, and packed into a rack-scale server called the Panjiu AL128 that links 128 chips with chip-to-chip latency under 150 nanoseconds.
Why it matters
Designing its own accelerators lets Alibaba own the full stack from chip to model, and reduces its dependence on foreign AI hardware at a time when access to that hardware is politically fraught. Owning every layer also lets the company tune the model and the silicon together, which is the thinking behind the long-running coding test. For enterprises in this region weighing where to run agent-style applications, a cloud provider that controls its own chips is pitching both cost and supply stability.
Availability
Alibaba announced the upgrades at the Alibaba Cloud Summit on 20 and 21 May 2026, positioning them as platform-level improvements for developers and enterprises on Alibaba Cloud. Some of the hardware rolls out in stages, and access for Malaysian customers will follow Alibaba Cloud's regional availability.