Deepseek Open‑Sources TileLang Toolkit for Huawei Ascend AI Chips

What happened, who, and when

On September 30, 2026, Deepseek announced via its official WeChat channel that it had completed a suite of open‑source programming tools for Huawei’s Ascend AI chips. The release, which includes libraries for computation and data movement, is backed by Huawei, which “fully supported” the effort. A key component of the bundle is TileLang, an open‑source language designed to simplify AI chip programming.

Concrete details of the release

• Software bundle: The open‑source package contains computation libraries, data‑transfer utilities, and a “supernode” configuration that links 128 Ascend 950 chips.
• TileLang: Originally created by researchers at Peking University, TileLang has been used internally by Deepseek for roughly a year. The company says the language offers a simpler programming model than Nvidia’s CUDA while delivering full hardware performance.
• Testing history: Deepseek first evaluated TileLang on older Nvidia GPUs before adopting it as the primary tool for its artificial general intelligence (AGI) work, according to The New York Times.
• Huawei’s role: Two weeks prior to Deepseek’s announcement, Huawei unveiled new AI processors and supernode systems slated for widespread model‑training use in the following year. The company also noted constraints on exporting chips abroad due to U.S. export controls, emphasizing a need for domestic software support.

Industry context and competitive landscape

• CUDA’s dominance: Nvidia’s lead stems not only from chip design but also from an estimated four million developers worldwide who build on its CUDA platform, creating a software moat that rivals like AMD have struggled to breach.
• Chinese AI ecosystem: Model developers such as Z.ai and Moonshot AI have outpaced domestic chip makers, prompting Huawei to accelerate its hardware‑software integration.
• Software as a differentiator: Research firm SemiAnalysis observed that while OpenAI’s Jalapeño chip outperformed Nvidia’s Blackwell on performance‑per‑watt in limited tests, Nvidia still leads on complex agent workloads measured by the AgentX benchmark. The analysis highlighted that Nvidia’s advantage lies in software that efficiently links many chips, a capability Huawei’s CANN stack shares only with CUDA for day‑one model support.
• Strategic implications: By open‑sourcing TileLang and associated libraries, Deepseek aims to create a universal, easy‑to‑program language that can foster a domestic AI‑chip software ecosystem, reducing reliance on foreign tools and helping China’s chip push advance.