AI Tools Compared: What Works in Hong Kong & China

Ask most people “which AI tool is best?” and you’ll get an argument about benchmarks. For a business in Hong Kong or Mainland China, that’s the wrong question. The ones that actually matter are: can you even use it where you operate, does it train on what you type, and where does your data end up? Get those wrong and the “best” model is a liability.

This is a practical comparison of the primary AI systems on exactly those points — access, privacy, and data location — for a Hong Kong or China business. It’s a companion to our practical AI advisory (the how to decide framework, with an interactive checker) and to can you use ChatGPT in Hong Kong?. One caveat up front: this is a mid-2026 snapshot, and AI providers change their terms and their geographic availability often — always check the live policy before you rely on it.

The one thing that matters most: who sees your data

Before any tool comparison, learn the single rule that prevents most mistakes:

• Consumer and free tiers — Western and Chinese — generally train on your inputs by default. ChatGPT Free/Plus, the consumer Gemini app, DeepSeek’s app, Kimi, Perplexity’s free tier — all may use what you type to improve their models (some with an opt-out, some weaker than others). Never put confidential, personal or regulated data into a consumer chatbot.
• Enterprise, business and API tiers from the Western providers generally do not train on your data by default — it’s a contractual commitment (OpenAI business/Enterprise/API, Anthropic commercial/API, Google Workspace, Microsoft 365 Copilot). That’s what makes them safe for real work.
• “Where the server is” is not the same as “who can compel the data.” A Chinese-rooted provider that stores data in Singapore is still subject to China’s PIPL and National Intelligence Law through its parent company. Legal control, not server geography, is the real privacy variable.

The comparison at a glance

Capabilities are left out of this table on purpose — they change monthly and every model is “good enough” at the common jobs; the columns below are the ones that don’t change as fast and decide more.

System	Hong Kong / China access	Trains on your inputs?	Where your data is held	Self-host?
ChatGPT / GPT (OpenAI)	Blocked first-party in both HK & China; reach it via Azure (Singapore) or Copilot	Business/Enterprise/API: no · Consumer: yes (opt-out)	At-rest options incl. Singapore; none in Greater China	No
Claude (Anthropic)	Blocked first-party in both HK & China; reach it via AWS Bedrock (Singapore)	Commercial/API: no · Consumer: yes (opt-out)	Bedrock Singapore region; none in Greater China	No
Gemini (Google)	Available in HK; unreliable in China; reach it via Vertex AI	Workspace/Enterprise: no · Consumer: yes unless disabled	EU/US (some Asia); none in Greater China	No
Microsoft 365 Copilot	Works in HK (enterprise)	No (prompts, responses, Graph data)	Microsoft data boundary; none in Greater China	No
DeepSeek	Available (Chinese app & API)	Consumer app: yes, default (weak opt-out)	Consumer app: in Mainland China	Yes — open weights
Kimi (Moonshot)	Available	Yes, by default	Int’l API: Singapore — but Chinese-jurisdiction parent	Partly
Qwen (Alibaba)	Available	Tier-dependent	China, or your own servers if self-hosted	Yes — Apache-2.0 (all sizes)
GLM (Zhipu) / MiniMax	Available	Tier-dependent	China, or your own servers	Yes — GLM MIT, MiniMax-M2 free for SMEs
Perplexity	Available in HK	Consumer: yes (opt-out) · Enterprise & Sonar API: no	North America; none in Greater China	No (answer engine)

The Western frontier models

These are the most capable general-purpose assistants — strong at writing, summarising, coding, vision and reasoning. The catch for this region is access, and the answer is almost always an enterprise route, not the consumer app.

ChatGPT / OpenAI

The most widely known, and very capable across the board. OpenAI blocks Hong Kong and Mainland China first-party — the consumer app and direct API aren’t officially available, and HK phone numbers are rejected. The practical route into GPT from this region is Microsoft 365 Copilot or Azure (see the governed-cloud section below). On data: OpenAI’s business, Enterprise and API tiers don’t train on your content by default; the consumer tiers do unless you opt out. Data-residency-at-rest is available in several regions including Singapore — but none in Greater China.

Claude / Anthropic

Widely rated for careful writing, long-document work and coding. Like OpenAI, Anthropic does not support Hong Kong or Mainland China first-party — Claude.ai and the direct API aren’t available, which is why some HK firms have had to switch staff off it. The governed route is AWS Bedrock (see below). Anthropic’s commercial and API usage is no-training by default; consumer tiers (since August 2025) train unless you opt out. Zero-data-retention is available on the API but must be arranged with Anthropic — it isn’t automatic.

Google Gemini

Capable and tightly integrated with Google Workspace; strong at long context and multimodal tasks. Gemini is reachable from Hong Kong (as of early 2026) and unreliable in Mainland China. Paid Workspace/Enterprise Gemini does not train on your data; the consumer app does unless you turn off Gemini Apps Activity. Workspace offers EU/US data regions (no Greater-China option). For developers, Vertex AI is the governed route.

Microsoft 365 Copilot

If you already run Microsoft 365, this is usually the lowest-friction safe choice in Hong Kong. Copilot and Copilot Chat do not use your prompts, responses or Microsoft 365 data to train foundation models — it’s covered by Microsoft’s data-protection addendum with Microsoft as your data processor, and the enterprise protection is included in Copilot Chat at no extra cost when you sign in with a work account.

The Chinese models

The Chinese AI field is excellent and moving fast — and, surprisingly, often more openly licensed than the Western open models. The privacy trade-off is the mirror image of the West: capable and accessible in-region, but the consumer services train on your inputs and your data sits under Chinese jurisdiction.

DeepSeek

Highly capable at reasoning and coding, and genuinely open-weight — you can download and self-host it (a major plus for data control). But the consumer app collects, processes and stores personal data in Mainland China by default, and trains on your inputs by default (the opt-out is weaker than the Western ones, and doesn’t delete history). For an SME, that makes the hosted app unsuitable for confidential data — while the open weights, run on your own infrastructure, are one of the better self-host options.

Kimi (Moonshot AI)

Strong long-context model. Its international API stores data on servers in Singapore — but Moonshot is a Beijing-founded company, so PIPL and the National Intelligence Law can reach that data regardless of where the server sits. It trains on inputs by default, with no-training available only via a negotiated enterprise agreement.

Qwen, GLM, MiniMax and ERNIE — the open-weight standouts

This is the counter-intuitive part. Several leading Chinese models ship under genuinely permissive open-source licences:

• Alibaba Qwen — Apache-2.0 across all sizes (the most useful open family for business self-hosting).
• Zhipu GLM-4.6 — MIT.
• MiniMax-M2 — a “modified MIT” that is free for SME commercial use (only firms above ~100M users or ~US$30M revenue owe an attribution notice).
• Baidu ERNIE 4.5 — reportedly open-sourced under Apache-2.0.

These are more permissive than Meta’s Llama (which carries a restricted community licence). Licences do vary by model and release, so check the specific model card — but the headline holds: if you want to self-host an open model, the strongest licence terms are largely Chinese.

Perplexity (and Comet)

Perplexity isn’t a chatbot so much as an AI answer engine — it searches the web and answers with citations, and now ships an AI browser (Comet) with agentic features. It’s available in Hong Kong. Privacy splits by tier: the consumer tiers train on your searches by default (opt-out in settings); the Enterprise tier and the Sonar API do not train and the Sonar API operates zero-data-retention (prompts aren’t stored) — a genuinely privacy-safe route. Data is processed in North America, with no Greater-China residency. The Sonar API is usage-priced (from about US$1 per million tokens plus a small per-search fee).

The route most people miss: frontier models through a governed cloud

Here’s the practical answer for Hong Kong and China that the “ChatGPT is blocked” headlines skip: you can use the Western frontier models compliantly through a cloud platform, even though the consumer apps are blocked.

• Claude → Amazon Bedrock, in the Asia Pacific (Singapore) region. Inference runs inside the AWS security boundary with zero Anthropic operator access, and data handling is governed by Bedrock’s terms.
• GPT → Microsoft Azure (Foundry / Azure OpenAI), with frontier models hosted in Singapore. You can run the management project from Hong Kong while the models are served from Singapore.
• Gemini → Google Vertex AI, similarly via Singapore.

In every case you get the model inside an enterprise security boundary, with no-training commitments and the ability to pin data processing to a chosen region — Singapore being the nearest to Greater China (there is still no Mainland-China or Hong-Kong region from any of them). For a regulated or cautious HK/China SME, this — not the consumer app — is usually the right way to use a frontier model.

How API access works, briefly

If you outgrow the chat apps, the API is how AI gets built into your own systems:

• You get an API key, and you pay per token (per unit of text in and out) — no monthly seat.
• The enterprise data commitments are stronger than the apps: no-training by default, and zero-data-retention available (gated/approved) from OpenAI and Anthropic.
• You can reach the same models through the hyperscalers — Azure OpenAI, AWS Bedrock, Google Vertex — which is what makes the governed, region-pinned route above possible, and what most businesses in this region should use.

Should you run a local (self-hosted) model?

“Run your own AI, on your own hardware, so your data never leaves” is appealing — especially for a China-data-residency, air-gapped, or offline need. Here’s the honest picture.

The models exist and are well-licensed. As above, Qwen (Apache-2.0), GLM (MIT), MiniMax and others are free to run commercially. So is Mistral (Apache-2.0) and, with more restrictions, Meta Llama and Google Gemma. DeepSeek’s open weights are a strong self-host option that sidesteps its hosted app’s China-data problem entirely.

The hardware is the real constraint (rough guide — verify for your model):

• A small model (~7–8B parameters), quantised, runs usefully on a good laptop or a modern Mac — fine for drafting, summarising and private note-taking.
• A mid-size model (~30–70B) needs a workstation GPU or a high-memory Apple-silicon machine — capable, but a real purchase and a noticeable step down in polish from the frontier hosted models.
• Frontier-class quality is impractical to self-host for almost any SME.

The tooling is mature: Ollama and LM Studio make running a model locally genuinely easy; vLLM serves models to a team. None of it removes the ongoing work — updates, security, scaling, and the quality gap to a hosted frontier model.

The honest verdict: self-host for control, not to save money. It’s worth it when you have a specific reason a hosted tool can’t meet — sensitive or regulated data that must not leave your premises, a China data-residency requirement, an air-gapped or offline environment. For most SMEs doing everyday work, a governed enterprise tier (or an open model hosted in your own cloud region) beats a local box on quality and far lower hassle. The right answer is a deliberate choice, not a default.

Choosing well

The pattern across this whole comparison is the one from our AI advisory approach: start from the job and the data, not the brand. Decide how sensitive the data is and where you operate, and the shortlist of safe tools picks itself — a consumer app for public-information tasks, an enterprise/governed-cloud route for confidential work, a self-hosted open model when control is non-negotiable.

PTS is vendor-neutral — we don’t sell a model — so we help you pick what fits your use case, jurisdiction and data, and then make it work safely, alongside your legal advisers and to our own ISO/IEC 27001 and 20000 disciplines. If you’d like a hand turning this into a sane, safe setup, talk to PTS.

Related reading: Practical AI advisory for Hong Kong & China · Can you use ChatGPT in Hong Kong? · Microsoft 365 in China · Hong Kong & China data laws