Private deployment depends on the actual model
Official DeepSeek and Qwen documentation describes downloadable weights and local execution, making selected models candidates for private AI. Not every product, API model or feature is downloadable: review each licence, restrictions, compute and application needs. This article provides general technical and business information, not legal advice or a legal compliance conclusion.
Public APIs: map the data route
Map the complete route of prompts, documents, responses and logs.
Further considerations
Identify services receiving prompts, uploaded documents, responses, logs and support material. Check retention, training use, subcontractors, processing locations, deletion and incident notification. Routing cannot override a third-party regional policy; choose officially available supply channels.
Hong Kong hosting: which layers are controlled?
Hong Kong hosting does not mean every external API, log or support process stays locally.
Further considerations
Hosting may cover applications, documents, indexes and suitable models. External models, monitoring or other APIs can still process data elsewhere. Map inference, storage, logs, backups and remote support separately, confirming locations and access.
On-site compute and maintenance ownership
Confirm capacity, dependencies and ownership of updates, monitoring and recovery.
Further considerations
Assess model memory, compute, capacity, networks and availability, with responsibility for updates, monitoring, recovery and incidents. Restricted or offline environments need dependency testing. Installing weights on a server alone does not complete implementation.
Sensitivity and retention
Use data for necessary purposes and manage retention, deletion and access records.
Further considerations
Classify personal data, business secrets, policies and public information. Send only fields necessary for approved purposes. Prompts, responses, indexes and logs may themselves contain sensitive data; set retention, deletion, access records and test-data rules.
Internal knowledge and permissions
Internal knowledge must respect departmental permissions; evaluate Scheme E/GenCMS where suitable.
Further considerations
Build sources around document owners, versions and departmental permissions. Retrieval and answers must respect visibility; test routine questions, unauthorised requests and unsupported answers with different roles. AskCore serves website customers; evaluate Scheme E/GenCMS for controlled internal knowledge and private AI.
PDPO and procurement questions
Document privacy and procurement questions for appropriate professional review.
Further considerations
Record purposes, necessity, notices, permitted uses, security, retention and vendor responsibilities for appropriate privacy/legal advisers to assess in the actual circumstances. Deployment location is one factor and does not automatically satisfy every requirement.
Scheme E / GenCMS as the implementation layer
Start with sources, permissions and one workflow, then plan integration and ongoing care.
Further considerations
iGears can organise sources, permissions, model routes, retrieval, application integration, tests and ongoing care, combining APIs, Hong Kong hosting and on-site delivery by sensitivity. Start with a defined workflow and acceptance conditions, then document responsibilities in the solution and written quotation.
Practical next step
References
- OpenAI — ChatGPT supported countries
- Anthropic — Supported Regions Policy
- Google — Gemini web app availability
- Microsoft — Microsoft 365 Copilot requirements
- DeepSeek — open-weight deployment documentation
- Qwen — official deployment documentation
- Alibaba Cloud — Model Studio regions and APIs
- PCPD — Artificial Intelligence guidance