October 21, 2026 — At its product launch, Baidu Dazi announced full adaptation to China's domestic IT stack and listed its agent software in Lenovo Kaitian's Tianxi Store, the Kylin Software Store and the UnionTech App Store. Together with Lenovo Kaitian, it also launched what the two companies call the first full-stack domestic agentic AI workstation.
The news did not dominate hardware forums, but it matters to anyone delivering IT projects to government or state-owned enterprises. For three years, "domestic AI workstation" usually meant domestic hardware plus a customer who had to install the software stack themselves. Delivery was slow, compatibility issues were common, and when something broke, vendors pointed at each other. A pre-integrated machine changes the economics of those projects.
1. What "full-stack" means
According to the launch materials, the workstation has four layers:
- Hardware: a Lenovo Kaitian chassis.
- Compute: domestic GPUs including Hygon BW100 and Tianshu Zhikai 100.
- Operating system: Kylin OS and UnionTech UOS.
- Software: Lenovo Tianxi AI Pro plus Baidu Dazi, with models and runtimes pre-installed.
2. Why it matters
The real bottleneck in domestic AI projects has not been the chip. It has been the length of the delivery chain:
| Layer | Old approach | What changes |
|---|---|---|
| Hardware | Domestic CPU/GPU boxes from multiple vendors | One Lenovo Kaitian platform and one warranty |
| OS | Customer installs Kylin or UOS; drivers are trial and error | Factory image, vendor-validated |
| Model runtime | Customer installs drivers and frameworks; version hell | Pre-installed models and runtimes |
| Agent software | Customer builds RAG and tool calls | Baidu Dazi plus Tianxi AI Pro out of the box |
| Support | Unclear who owns the problem | Joint vendor support with a defined boundary |
For integrators and end customers, that integration is worth more than any single specification. The most expensive part of a domestic IT project is rarely the hardware; it is the weeks spent on compatibility testing and the meetings spent assigning blame.
3. Who it is for
- Government and state-owned enterprises: data cannot leave the premises, and the stack must be domestic. Local knowledge bases, document drafting and meeting minutes are the obvious first workloads.
- Education and healthcare: campus or hospital knowledge bases, structured records and research data, where compliance is sensitive.
- Manufacturing and energy: equipment-manual Q&A, first-pass image inspection and work-order assistance, usually on an isolated plant network.
- Mid-sized companies that do not want cloud lock-in: keep the agent and the data in their own server room.
It is not the right choice if your workload depends on custom CUDA operators, the newest training frameworks or maximum per-card performance. This is a compliance-first, out-of-the-box platform, not a performance-first one.
4. Five questions to ask before you buy
- Which GPU, and how much memory? Hygon BW100 and Tianshu Zhikai 100 are different architectures. Check memory capacity, precision support and the official model compatibility list, not just the phrase "domestic GPU supported."
- Which models are pre-installed? Parameter size, quantization, and whether you can import your own models from Hugging Face or ModelScope.
- Which OS version? Kylin and UOS versions differ in Active Directory, printing and browser support. Confirm integration with your existing domain, OA and business systems.
- What can the agent actually do? Which local tools can Baidu Dazi call? How are permissions enforced? Is there a private knowledge base and an audit log?
- What does support look like? How long is the warranty? Is a failed GPU replaced as a card or as a whole unit? Is there a service partner in Guangzhou?
5. Three traps
- "Domestic GPU supported" does not mean "your model runs." Operator coverage, precision and memory differ by card. Check your actual model against the vendor's compatibility list.
- Ignore OS and domain compatibility at your peril. Kylin and UOS versions vary widely in AD domain join, print services and browser engines. Discovering this at acceptance is expensive.
- Ignore spare-parts lead time. Domestic GPU repair cycles are often longer than established brands. Build a spare unit or a parts pool into the project, or one failed card can stop the business for two weeks.
6. A four-step rollout
- Pick one scenario. Choose a high-frequency, low-risk, measurable task such as internal policy Q&A. Do not start with a core business process.
- Run a proof of concept. Use real data for two to four weeks. Track accuracy, response time and the human-takeover rate against your current process.
- Buy a small batch. Start with one or two units. Prove networking, permissions, backup and auditing before scaling.
- Then expand. Write down model updates, knowledge-base maintenance and incident handling, and define who owns what between the vendor and the integrator.
7. What ACCPC can do
ACCPC does not manufacture domestic hardware. We do three things around it: compare platforms (Hygon, Tianshu, Ascend) against your actual software list, integrate the workstation into your existing network, domain, backup and monitoring, and maintain it after deployment. If you are preparing a domestic AI project or an acceptance review, send us the requirements and we will start with a feasibility check.