Huawei has released a couple of new agentic AI solutions for enterprises during the Huawei Cloud INSPIRE 2026 event at the West Bund International Convention and Exhibition Center in Shanghai.
The new product lineup empowers the agentic AI infrastructure in the industry and includes the following solutions:
- Agentic Infra unified infrastructure for general & AI workloads
- New-generation model training and inference platform
- An enterprise-grade agent platform
The new Agentic Infrastructure features an efficient token factory, continuous learning, unified general and AI compute scheduling, and secure autonomy. Here are the products that make the Agentic Infra recognizable.
AICS – Its ultra-high bandwidth UnifiedBus (UB) network supports clusters with more than 100,000 cards. Thus, providing a total computing power of up to 200 EFLOPS. AICS reduces token generation latency to less than 10 million seconds, with a throughput of 5 million tokens per second across 1,000 cards. Its online service availability tops 99.5%.
AMS – the Agentic Memory Storage system uses NPU passthrough to Context Memory Storage (CMS) hardware to create a PB-scale memory space. The solution supports tiered KV-cache pooling, reducing inference costs while enabling multi-day long-running tasks, resolving the memory bottleneck of agents, and improving continuous learning for agents.
CCE Volcano unified general and AI scheduling engine: It achieved unified general-purpose and AI workload scheduling through “shared training-inference pooling + fragmentation consolidation,” to improve resource utilization by over 30%.
AgentSphere – It provides a secure and autonomous agent runtime environment, providing a secure autonomous foundation with seamless and proactive intent protection. Leveraging utla-lightweight sandbox technology, achieving fast startup with 100 milliseconds and the ability to batch-create hundreds of thousands of instances per minute, allowing agents to scale securely and efficiently on the cloud.
ModelArts:
This is a new model training and inference platform that has four core capabilities:
- Reinforcement Learning as a Service (RLaaS)
- Confidential inference
- Model routing
- Model matrix
MaaS model routing supports three policies – experience-first, efficiency-first, and balanced mode, and dynamically routes each request to the optimal model based on its characteristics.
As of today, the company has released over 15 state-of-the-art (SOTA) model services with a model scheduling accuracy of over 95% and an average reduction of 20% in calling costs. The enterprise-level RLaaS service allows reinforcement learning as a core capability that can be invoked by every enterprise.
“It allows users to create tasks in just one minute, achieve end-to-end visualization, and ensure consistency between training and inference. This enables large models to be applied to more specific scenarios and become smarter with each use,” wrote Huawei.
Read everything about other AI agent solutions at Huawei’s official website.
The post Huawei launches new agentic AI solutions for enterprises appeared first on Huawei Central.
