T-Head Unveils Zhenwu V900 AI Chip and Expands Open-Source Software Stack

T-Head presents the AI chip as the engine and the software stack as the gearbox, transmission system and driving system. T-Head SAIL is a CUDA-like stack built around the Zhenwu AI chip. It connects upper-layer AI frameworks such as PyTorch to Zhenwu hardware and handles model migration, compilation and execution, computing acceleration, and development debugging. In July at WAIC, T-Head said it had started open-sourcing T-Head SAIL and provided downloads for SDKs, drivers, performance analysis and debugging tools, and technical documentation. At the Yunqi Conference, the team disclosed further progress. Open-source projects now include PyTorch-for-sail for framework adaptation, sailify for source migration, Triton-for-sail for operator development, and acceleration projects such as DeepGEMM-for-sail and FlashAttention-for-sail.

Zhenwu AI chips have served more than 650 customers in over 20 industries, and leading companies including Ant Group, Xiaohongshu, and XPeng have begun using T-Head SAIL. According to the report, Xiaohongshu developed a model migration and operator optimization Agent based on SAIL open-source code to speed up deployment of generative recommendation models on Zhenwu. T-Head supplies base software capabilities, while customers write their business experience into tools and continue migration and optimization work, the company said.

XPeng used SAIL to migrate business models that previously ran on GPU platforms to Zhenwu cloud clusters for intelligent driving model training. After migration, the team used SAIL performance analysis tools to locate training bottlenecks and optimize around operators and frameworks. SAIL supports development methods including C++, Triton, and TileLang, aiming to reduce relearning burdens and let teams focus on models and business. Ant Group's inference service team has completed inference adaptation for major frontier models on Zhenwu 810E and M890 based on SAIL. Work continues on quantization, prefill-decode separation, expert parallel load balancing, and sparse attention integration. Xiaohongshu's Agent is another example of customers turning migration and optimization into their own tools.

Lu Shenghua, senior director of T-Head's semiconductor software ecosystem, said in a public speech that 'the biggest demand, from a business perspective, is how high the migration cost is.' Many customers have software that has run online for years, with repeatedly revised code, familiar tools, and internal tuning experience. Switching chips means those accumulated assets must be tested again without interrupting business. Framework adaptation, source migration, and operator development tools are meant to preserve as much of that work as possible. Migration alone is not enough. If efficiency falls to 30% to 40% of the original after switching chips, businesses will find it hard to accept even if the model runs. Open-source acceleration projects such as DeepGEMM-for-sail and FlashAttention-for-sail allow developers to see implementations and modify them for their own models and workloads, rather than sending every problem back to the chip vendor.

Developers also expect new models to run on Day 0. Model updates now come every few weeks, and slow adaptation delays business teams from trying new technology. T-Head is providing model resources adapted for Zhenwu. As of September 2026, it offered 39 quantized models on ModelScope covering Qwen, DeepSeek, and Kimi series, with cumulative downloads exceeding 348,000, according to the speech. TensorFlow and JAX adaptation versions, an in-house inference engine, and communication components and debugging monitoring tools including PCCL and DeepEP-for-sail are still being open-sourced.

T-Head's chip effort has moved from the Hanguang 800 in 2019 to the Yitian 710 in 2021 and now the Zhenwu series. One continuing path is to design chips from business needs, validate them first inside Alibaba, and then serve external customers through cloud services. Lu said external customers often first ask whether Alibaba has used the chip itself. Taobao, search, recommendation, and advertising businesses played an early production validation role. With the Zhenwu M890, the team has completed two full generations of chip delivery. In Lu's view, software maturity and customer demand for secondary development together prompted the opening at this time. More than 650 customers means the product is being used, but also that the vendor faces increasingly varied problems. Automakers train models, internet companies optimize recommendation, and model teams try new algorithms. Their computing patterns and performance bottlenecks differ. It becomes unsustainable for the chip vendor to handle all business-specific development on one base software stack.

T-Head also faces long-term maintenance costs. Chip vendors adapting to mainstream frameworks usually modify some code. Maintaining an independent version means resynchronizing code, handling differences, and testing every time upstream software updates. Submitting adaptations and optimizations back to upstream communities such as PyTorch, vLLM, and Triton is intended to let Zhenwu support evolve with mainstream software and reduce repeated maintenance. After T-Head SAIL was open-sourced, Alibaba internal and external customers have submitted code contributions, according to the report. Feedback includes requests to contribute projects, more hardware architecture information for developing custom operators, and faster support for new models and frameworks. Zhenwu architecture information has been made public, and community governance and contribution rules are still being improved. The team is promoting decoupled releases of software components so different projects can update separately, and is gradually submitting Zhenwu adaptation results upstream.

At the Yunqi Conference, Wu said Alibaba would invest long term in AI models, AI chips, and AI cloud. In February 2025, Alibaba announced a plan to invest at least 380 billion yuan over three years in cloud and AI infrastructure, and later said it would further increase investment under the plan. T-Head is working with the Qwen team to explore using models to generate high-performance operators. Generating functionally correct code is relatively easy, but making it fully use chip performance requires the model to understand the underlying architecture. Qwen's needs cannot cover all industries. Xiaohongshu's generative recommendation and XPeng's intelligent driving training each have their own models, data, and optimization goals. The further opening of T-Head SAIL is meant to let such teams participate in software development connecting models and hardware, turning their understanding of business into concrete implementations. SAIL stands for Seed of AI Library. T-Head says each line of open-source code is a seed.