Alibaba Unveils New AI Chip, Flagship Model, and Rebuilt Cloud Stack AI for Agentic Era

Alibaba on Wednesday launched its most aggressive AI push yet, unveiling a new flagship large language model, a homegrown AI chip that triples the performance of its predecessor.

Qwen3.7-Max debuts for agentic workloads; new Zhenwu M890 triples the performance of its predecessor
Model and application services platform ARR to surpass RMB 10 billion (approximately US$1.4 billion) in the June quarter and reach RMB 30 billion (US$4.1 billion) by year-end

Alibaba on Wednesday launched its most aggressive AI push yet, unveiling a new flagship large language model, a homegrown AI chip that triples the performance of its predecessor, and a rebuilt cloud platform designed from the ground up for autonomous AI agents. The announcements came at the Alibaba Cloud Summit in Hangzhou and span every layer of the company’s technology stack, from silicon to software.

Qwen 3.7-Max: A Model That Works, Not Just Answers

Qwen3.7-Max, Alibaba’s latest large language model, is built for agentic workloads. The model is engineered for sustained, multi-step operations rather than single-turn responses.

The model’s defining capability is endurance. In an internal benchmark shared at the summit, Qwen 3.7-Max was given a task brief and placed on a Zhenwu M890 chip it had never encountered in training. Working without human intervention, the model ran for 35 consecutive hours, executed more than 1,000 tool calls, and delivered a production-grade AI computing kernel that outperformed the chip manufacturer’s official version by tenfold.

The result illustrates a shift that Alibaba believes defines the agentic era: a model that completes complex engineering tasks on its own. Qwen3.7-Max also handles multi-file software projects, orchestrates multi-agent office workflows, and is optimized for agent harnesses including OpenClaw, Hermes Agent, Claude Code, Qwen Paw and Qoder.

A Cloud Rebuilt for Agents

Alibaba Cloud also launched the Panjiu AL128 Supernode Server, a rack-scale system that packs 128 AI accelerators into a single unit and delivers petabyte-per second (PB/s) internal bandwidth. The configuration is designed specifically for the concurrency patterns that agents generate: unpredictable, high-frequency bursts of inference requests that overwhelm conventional compute clusters.

The server is now available through Alibaba’s model service platform, Bailian (also known as Model Studio outside China). The platform has introduced Agentic RL, a reinforcement learning mechanism that refines models continuously based on actual agent task outcomes. It also features built-in safety governance guardrails that keep autonomous agents operating within defined boundaries.

Homegrown Silicon: T-Head Steps Up

T-Head, Alibaba’s chip design subsidiary, formally debuted the Zhenwu M890, its most powerful AI processor to date. The chip delivers three times the performance of its predecessor, Zhenwu 810E, and carries 144 gigabytes (GB) of on-chip memory alongside 800 GB per second of inter-chip bandwidth. It natively supports precision formats from FP32 (32-bit floating-point) down to FP4 (4-bit floating-point), allowing a single device to handle both high-accuracy model training and the rapid, low-cost inference that agent workloads demand.

T-Head paired the Zhenwu M890 with a new networking chip, ICN Switch 1.0, which delivers up to 25.6 terabits per second (Tbps) of aggregate bandwidth and enables congestion-free communication across clusters of 64 accelerators. The company also launched T-Head SAIL™, a software stack designed to extract maximum performance from its proprietary hardware. Together, the chips underpin the Panjiu AL128 server.

T-Head has shipped more than 560,000 Zhenwu chips to date. More than 400 external customers across 20 industries have deployed them, including leading automakers and financial services companies.

The announcements signal a strategic repositioning. During its latest earnings announcement, Alibaba CEO Eddie Wu shared that the company expects its model and application services platform annual recurring revenue (ARR) to surpass RMB 10 billion (approximately US$1.4 billion) in the June quarter and projected that figure to reach RMB 30 billion (US$4.1 billion) by year-end. He forecasted that AI-related product revenue will surpass conventional cloud compute sales as Cloud Intelligence Group’s largest revenue line in approximately one year.

This article was originally published on Alizila written by Crystal Liu.

Community

Alibaba Unveils New AI Chip, Flagship Model, and Rebuilt Cloud Stack AI for Agentic Era

Qwen 3.7-Max: A Model That Works, Not Just Answers

A Cloud Rebuilt for Agents

Homegrown Silicon: T-Head Steps Up

Read previous post:

Alibaba Cloud Community

You may also like

Comments

Alibaba Cloud Community

Related Products

Alibaba Cloud Model Studio

Qwen

Alibaba Cloud for Generative AI

AgentBay