Alibaba unveils Zhenwu V900 chip and opens T-Head SAIL stack

China's cloud giants are moving from shipping chips to building ecosystems around them. That is how qbitai (量子位) framed the remarks of Alibaba Group CEO Wu Yongming (吴泳铭) at the Yunqi conference on September 22. Presenting the new AI chip Zhenwu V900, he called it the most powerful computing chip in China. He said it should be three times more efficient than the earlier M890.

A day later, T-Head announced the next stage of opening the T-Head SAIL software stack, the counterpart of CUDA for Zhenwu chips. The open-source release includes the PyTorch-for-sail adaptation, the sailify code migration tool, the Triton-for-sail environment for writing operators, and the DeepGEMM-for-sail and FlashAttention-for-sail libraries. The first components appeared in July at the WAIC fair. T-Head released the SDK, drivers, and tools for performance analysis and debugging then.

The problem it is meant to solve

The argument sounds familiar to anyone who has moved production onto a different hardware platform: migration cost matters most. Lu Shenghua, a senior director at T-Head, said plainly that the first thing customers ask is how much they will have to rewrite. A second problem shows up once a model is running. If performance drops to 30 to 40 percent of the previous level after the hardware change, nobody will accept such a migration. The third is pace. Models come out every few weeks. Customers expect a new model to run at least on its launch day.

Opening the code of the acceleration libraries shifts the balance of power for another reason. A customer that wants to write its own operators sometimes has to reveal details of its algorithm to the chip supplier. It does not want to do that. Access to the architecture and the code lets it do the work in-house. T-Head also describes how changes get submitted: a proposal, a code modification, automated tests, and a review by maintainers, then a release. According to the company, the first contributions came both from inside Alibaba and from external customers.

Who already uses it

Figures from the presentation put Zhenwu chips at more than 650 customers across more than 20 industries. Ant Group has finished adapting its most important models for inference on Zhenwu 810E and M890 chips. It is working on quantization and expert parallelism. XPeng moved training of its autonomous driving models onto a Zhenwu cluster. It uses profiling tools to find bottlenecks. Xiaohongshu went further. It built its own agent for migrating and optimizing operators on top of the open SAIL code. This sped up deployment of generative recommendation models.

The ecosystem also includes a model catalog. By September 2026, 39 quantized models had been published on the ModelScope platform. They cover the Qwen, DeepSeek, and Kimi families. The models were downloaded more than 348 000 times in total. The product line shows the direction: from the Hanguang 800 chip of 2019, through Yitian 710 of 2021, to today's Zhenwu series. The pressure toward openness is not altruism. Without portable code, foreign and domestic customers have no reason to switch platforms.