--- title: "Alibaba Cloud Summit: Launch of \"Zhenwu M890\" AI Chip and Debut of Qwen3.7-Max Flagship Model" type: "News" locale: "en" url: "https://longbridge.com/en/news/287015135.md" description: "At the Yunfeng Summit, Alibaba unveiled its self-developed Zhenwu M890 chip, featuring tripled performance and 144GB of built-in video memory, alongside a 128-card super-node server, with plans to launch the V900 in 2027. The flagship model Qwen3.7-Max was also released, boasting long-range agent capabilities that enable over 1,000 autonomous tool calls across 35 hours, ranking first among domestic models in blind tests" datetime: "2026-05-20T06:19:10.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/287015135.md) - [en](https://longbridge.com/en/news/287015135.md) - [zh-HK](https://longbridge.com/zh-HK/news/287015135.md) --- # Alibaba Cloud Summit: Launch of "Zhenwu M890" AI Chip and Debut of Qwen3.7-Max Flagship Model Alibaba Group concentrated its AI technology reveal at its annual Yunfeng Summit on Wednesday, launching the new generation of self-developed AI chip Zhenwu M890, a 128-card super-node server, and the flagship large model Qwen3.7-Max. This comprehensive display highlighted its latest progress in the full-stack system of "chip-cloud-model-inference." The Zhenwu M890, developed by T-Head, Alibaba's semiconductor subsidiary, delivers three times the performance of its predecessor, the Zhenwu 810E. Designed specifically for the era of AI agents, the chip handles the high-intensity memory and communication loads required for large-scale concurrent inference and long-chain tasks. Alibaba simultaneously launched a 128-card super-node server based on this chip, which is now available to Chinese enterprise customers via the Alibaba Cloud Bailian platform. The newly released flagship model, Qwen3.7-Max, ranks first among domestic models in the global blind test leaderboard for large models conducted by the third-party organization Arena, with performance approaching the strongest versions of GPT, Claude, and Gemini. Alibaba stated that the model can autonomously complete over 1,000 tool calls within a continuous 35-hour period, demonstrating persistent and stable long-cycle execution capabilities, making it one of the most representative foundational models for long-range agents currently available. These launches collectively present Alibaba's investment direction in AI infrastructure. Last year, Alibaba committed to investing over RMB 380 billion (approximately USD 53 billion) in cloud and AI infrastructure over the next three years, and this summit marks a phased fulfillment of this strategic layout. ## Self-Developed Chip Roadmap: Three Generations in Three Years, Rise of Domestic Chips The Zhenwu M890 features 144GB of built-in video memory, with inter-chip interconnect bandwidth reaching 800GB/s. It natively supports various data precisions from FP32 to FP4, covering full-scenario needs for high-precision training, low-precision, and ultra-low-precision inference. Paired with the self-developed ICN Switch 1.0 chip, it enables full-bandwidth interconnection for 64 cards, with 128 chips forming a single computing unit, reducing communication latency to the hundred-nanosecond level. Alibaba also announced its multi-year chip roadmap: **It plans to launch the next-generation chip, the V900, in the third quarter of 2027, with performance approximately three times that of the M890; the J900 will follow in the third quarter of 2028.** Gao Hui, Vice President of T-Head Semiconductor, revealed that cumulative shipments from T-Head have reached 560,000 units, a significant increase from the previously disclosed 450,000 units. Its customer base covers more than 400 enterprises across 20 industries, including automotive manufacturing and financial services. According to Reuters, T-Head has previously been spun off from Alibaba Group and is preparing for an independent IPO. ## Qwen3.7-Max: Benchmarking Long-Range Agent Capabilities Alibaba positions Qwen3.7-Max as the "flagship model for the agent era," with its core differentiation lying in its long-chain autonomous execution capability. In a test published by the company, Qwen3.7-Max optimized a production-grade AI inference kernel on the T-Head Zhenwu M890 chip platform in a fully autonomous manner. The process took approximately 35 hours and completed 1,158 tool calls, ultimately achieving a geometric mean speedup of 10x compared to the reference implementation. Comparative tests showed that for the same task, GLM 5.1 achieved a 7.3x speedup, Kimi K2.6 reached 5.0x, DeepSeek V4 Pro hit 3.3x, while Qwen3.6-Plus managed only 1.1x. Qwen3.7-Max's performance on several mainstream benchmarks also provides quantitative support for its market competitiveness: In reasoning ability, it scored 92.4 on GPQA Diamond, higher than Opus-4.6's 91.3; in programming agents, it scored 80.4 on SWE-Verified, basically on par with industry-leading levels; and on the general agent benchmark MCP-Mark, it scored 60.8, surpassing GLM-5.1's 57.5. Alibaba emphasized that these evaluations were based on entirely new out-of-domain environments not seen during model training, verifying true capability generalization rather than benchmark-specific optimization. Qwen3.7-Max will soon be available via API through the Alibaba Cloud Bailian platform, supporting mainstream agent frameworks such as Claude Code, OpenClaw, and Qwen Code. ## Behind the Full-Stack Layout: Soaring Computing Demand Parallel with Underlying Architecture Dividends Behind Alibaba's concentrated launches lie two parallel driving forces: first, the exponential growth in AI computing demand driven by the accelerated penetration of enterprise-level Agent applications; second, as cloud computing enters the era of hard-core technology, major tech firms are strategically reshaping their focus towards autonomous control of underlying technologies and supply chain resilience. On the chip front, the debut of the Zhenwu M890 enables Alibaba to provide vertically integrated solutions ranging from chips to cloud services for enterprise customers. T-Head's independent IPO plan may further unlock its commercial potential, allowing its service scope to extend beyond the Alibaba ecosystem itself. On the model front, the core scenarios targeted by Qwen3.7-Max—long-range autonomous task execution, cross-framework compatibility, and enterprise workflow automation—align closely with the mainstream demands of current enterprise AI deployments and directly correspond to the customer structure of Alibaba Cloud in the domestic cloud computing market. Alibaba's committed three-year investment plan of RMB 380 billion last year is gradually being implemented through chip iterations, model upgrades, and server system releases, reflecting its strategic orientation to treat AI infrastructure as the core engine for medium-to-long-term growth. ### Related Stocks - [09988.HK](https://longbridge.com/en/quote/09988.HK.md) - [KBAB.US](https://longbridge.com/en/quote/KBAB.US.md) - [BABO.US](https://longbridge.com/en/quote/BABO.US.md) - [BABX.US](https://longbridge.com/en/quote/BABX.US.md) - [SOXL.US](https://longbridge.com/en/quote/SOXL.US.md) ## Related News & Research - [Alibaba unveils new Qwen model, custom chips in bid to become China’s AI factory](https://longbridge.com/en/news/287044112.md) - [Andrew Sobko’s Argentum AI Expands Focus on Institutional AI Infrastructure Financing](https://longbridge.com/en/news/286140551.md) - [Cerebras's stock set for blast off, as early indications point to a near doubling](https://longbridge.com/en/news/286434594.md) - [Neurovia AI CTO at ISNR: Solving the AI Data Cost Dilemma and Unleashing Infrastructure Capacity | AIIO Stock News](https://longbridge.com/en/news/287049307.md) - [AI face is taking over — and driving plastic surgeons crazy](https://longbridge.com/en/news/286641783.md)