Skip to content

Technical Highlights ​

Document Information

Version: v1.0 Updated: 2026-07-13

Core Highlights ​

1) Unified Management of Heterogeneous Platforms and Compute ​

  • Multi-platform compute management: Manages configured and validated cloud resources together with listed local accelerators from NVIDIA, Huawei Ascend, Enflame, Biren, Hygon, and others. Huawei Cloud access is currently temporarily unsupported.
  • Resource-aware scheduling: Selects available compute based on regions, resource pools, specifications, templates, quotas, and authorization. Actual scheduling depends on the deployed version and underlying resources.
  • Unified operations entry points: Cluster, node, device, job, and monitoring pages expose the state of authorized resources.

2) Unified Management of Heterogeneous Models ​

  • Multi-source model management: Supports documented publication paths for single models, BYOK endpoints, and aggregate models. Exact sources and fields depend on the model page and deployed version.
  • Aggregate-model routing: Model providers can create aggregate models from eligible member models and select routing strategies available in the current version.
  • Publishing and review workflow: Model providers maintain models and submit reviews; operators process model and app reviews; end users consume authorized models.
  • Model calling entry points: Users can start from Model Marketplace, Playground, and the call example shown for a model. Protocol and model capability support must be confirmed by version.

3) Enterprise-grade Security and Isolation ​

  • Tenants, roles, and resource authorization scopes separate visible menus, operations, and data.
  • Review records, call logs, usage, and metering pages provide evidence for governance and troubleshooting.
  • Private and intranet deployment paths are available; data boundaries still depend on the actual network, storage, endpoint, and operations design.

4) Observability and Operability ​

  • Deployment status, monitoring, call logs, and call analytics pages expose runtime information.
  • Role scope, time range, resource state, and source logs can be combined to investigate deployment or call issues.

5) A Delivery System Built for Production Adoption ​

  • Requirement surveys, environment surveys, installation prechecks, scenario guides, and user manuals provide delivery entry points.
  • PoC-to-production delivery can confirm resources, permissions, compatibility, capacity, and acceptance results in stages.
  • Reuse of existing enterprise infrastructure requires validation against the installation requirements and project environment.

Differentiated Value ​

DimensionAGIOneTypical DIY Stack
Delivery efficiencyStandard path across surveys, installation, scenarios, and manualsMulti-component workflows and documentation must be organized independently
Operational complexityResource, model, billing, settings, and operations entry points across five product modulesMultiple systems and ownership boundaries are maintained separately
Access governanceTenants, roles, and authorization scopes control visible contentA permission system must be designed and integrated separately
Compute managementManages validated cloud resources and heterogeneous acceleratorsResources are commonly managed in separate systems
Model aggregationProviders can configure aggregate models and available routing strategiesEndpoints and routing logic are maintained model by model

Current status: Huawei Cloud access is temporarily unsupported; RAG and Function Calling are planned. Review the Support Matrix for other boundaries and the User Manual for operations.