DeepSeek for WhatsApp Business and Telegram Bots is a practical way to add AI replies, lead qualification, knowledge-base answers, and customer support automation to two of the world’s most used messaging channels. The core idea is simple: receive a message from WhatsApp or Telegram, send it to the DeepSeek API through a secure backend or workflow tool, apply business rules, and return a useful answer.
The implementation details are different. WhatsApp Business has strict rules around opt-in, approved message templates, the 24-hour customer service window, automation, and escalation. Telegram bots are usually easier to launch, but they still require secure token handling, webhook or polling setup, abuse controls, and careful data protection. DeepSeek’s current API supports OpenAI- and Anthropic-compatible formats, with deepseek-v4-flash and deepseek-v4-pro listed as current model options in the official docs.
Last reviewed: June 12, 2026. DeepSeek model names, pricing, WhatsApp Business rules, Telegram Bot API behavior, and integration options may change; always verify the official documentation before production deployment.
Table of Contents
What Is DeepSeek for WhatsApp Business and Telegram Bots?
DeepSeek for WhatsApp Business and Telegram Bots means using DeepSeek as the AI engine behind automated conversations on WhatsApp Business and Telegram. DeepSeek is not automatically built into either messaging channel. You connect it through the DeepSeek API, a workflow automation platform, or your own backend.
A typical bot receives user messages, classifies intent, retrieves relevant business data, asks DeepSeek to draft a response, applies safety and business rules, and sends the final answer back to the user. This can support customer service, order updates, internal team assistants, multilingual FAQ automation, lead qualification, appointment booking, and knowledge-base search.
The DeepSeek API is currently documented as compatible with OpenAI and Anthropic API formats. In the official quick-start docs, DeepSeek lists https://api.deepseek.com as the OpenAI-format base URL and shows deepseek-v4-flash and deepseek-v4-pro as current model values. Legacy names such as deepseek-chat and deepseek-reasoner are documented as deprecated from July 24, 2026, with compatibility mapping to deepseek-v4-flash.
Why Use DeepSeek for Messaging Bots?
DeepSeek is attractive for messaging automation because many WhatsApp and Telegram bots need fast, high-volume, conversational responses rather than one-off long-form reasoning. For everyday support, a lower-cost model can answer common questions, summarize policies, classify intent, and draft responses. For complex escalations, a stronger model or thinking mode can be reserved for difficult cases.
As of June 2026, DeepSeek’s pricing page lists token-based billing per 1M tokens and supports context caching, JSON output, tool calls, and a 1M context length for both deepseek-v4-flash and deepseek-v4-pro. DeepSeek also states that prices may vary and recommends checking the pricing page regularly before topping up or forecasting usage.
For businesses, the strongest reasons to use DeepSeek in messaging bots are:
- It can reduce repetitive support workload.
- It can support multilingual replies.
- It can connect to CRMs, order systems, help desks, and internal databases.
- It can run through no-code tools for prototypes.
- It can be deployed through custom code for stronger security, routing, and scale.
- It can use JSON output and tool calls for structured workflows such as lead scoring, appointment booking, and ticket creation.
WhatsApp Business vs Telegram Bots: Which One Should You Choose?
WhatsApp Business and Telegram can both work with DeepSeek, but they are not interchangeable. WhatsApp is usually better for customer-facing business communication where users already expect brand support. Telegram is often better for communities, internal teams, technical users, and fast experimentation.
| Factor | WhatsApp Business | Telegram Bot |
|---|---|---|
| Best use case | Customer support, ecommerce, sales follow-up, appointment reminders | Communities, internal assistants, developer tools, alerts, lightweight AI bots |
| Setup complexity | Higher | Lower |
| Main setup path | Meta Business account, WhatsApp Business Platform, Cloud API, phone number, webhook | BotFather, bot token, webhook or getUpdates |
| Message rules | Strict: opt-in, templates, 24-hour customer service window | More flexible bot interaction model |
| Outbound messaging | Usually requires approved templates outside the service window | Bot can message users who have started the bot, subject to Telegram rules and limits |
| Business suitability | Strong for official customer communication | Strong for communities and internal workflows |
| Community suitability | Limited compared with Telegram | Strong for groups, channels, and communities |
| API flexibility | Powerful but policy-heavy | Developer-friendly and quick to test |
| Compliance burden | Higher | Lower, but still requires privacy and security controls |
| Human handoff | Clear escalation path required; live agent handoff recommended | Recommended for support scenarios |
Meta’s WhatsApp Business policy says businesses may initiate conversations only with approved message templates, may reply without a template within 24 hours of the user’s last message, and may use automation during the 24-hour window only with prompt, clear, and direct escalation paths such as in-chat human transfer, phone, email, web support, store visits, or a support form.
Telegram’s Bot API is simpler technically: the official docs state that bots receive updates through either getUpdates or webhooks, that these methods are mutually exclusive, that updates are JSON-serialized, and that incoming updates are not stored longer than 24 hours.
High-Level Architecture
A production-ready DeepSeek messaging bot should not simply forward every message to AI. It needs a controlled architecture.
Basic flow:
User message → WhatsApp/Telegram → Webhook or polling → Backend/orchestration layer → DeepSeek API → Business logic/CRM/knowledge base → Response → User
A more complete system may include:
- A webhook endpoint for WhatsApp or Telegram.
- An orchestration layer such as Node.js, FastAPI, n8n, Make, or another automation platform.
- A prompt and policy layer.
- A CRM or ticketing system.
- A product, order, or appointment database.
- A vector database or knowledge base for RAG.
- A human handoff route.
- A moderation and abuse-prevention layer.
- A queue and retry system.
- Logging, analytics, and alerting.
For WhatsApp, the orchestration layer must understand whether the message is inside the customer service window and whether a template is required. For Telegram, it must handle update types safely, ignore unsupported messages, rate-limit abusive users, and avoid exposing the bot token.
How to Build a DeepSeek WhatsApp Business Bot
A DeepSeek WhatsApp Business bot is best treated as a customer-support workflow, not an unrestricted general-purpose AI assistant. That distinction matters because WhatsApp Business rules are stricter than Telegram’s, and the policy landscape around general-purpose AI assistants on WhatsApp has been changing.
In June 2026, Reuters reported that EU regulators ordered Meta to allow rival AI chatbots free access to WhatsApp during an antitrust probe, while Meta said it would appeal. The case specifically involves access for rival AI chatbots and WhatsApp’s Business API, so companies building AI products on WhatsApp should verify current Meta terms before launch.
Step 1: Prepare the WhatsApp Business setup
You will usually need:
- A Meta Business account.
- A WhatsApp Business Account.
- A registered business phone number.
- Access to the WhatsApp Business Platform or Cloud API.
- A webhook endpoint to receive inbound messages.
- Message templates for outbound or re-engagement messages.
- A privacy policy and consent/opt-in process.
- A clear escalation path to a human or support channel.
Meta’s policy requires necessary notices, permissions, and consents to collect, use, and share user content and information. It also restricts sensitive identifiers such as full payment card numbers, financial account numbers, and personal ID numbers.
Step 2: Design the WhatsApp conversation rules
Before connecting DeepSeek, define what the bot is allowed to answer. For example:
- Product availability.
- Shipping status.
- Return policy.
- Appointment scheduling.
- Lead qualification.
- Store hours.
- Order lookup after user verification.
- Handoff to a human agent.
Avoid letting the bot improvise legal, medical, financial, identity, or account-security instructions unless your business is properly compliant and reviewed by qualified professionals.
Step 3: Connect WhatsApp messages to DeepSeek
A practical WhatsApp workflow looks like this:
- Customer opts in to WhatsApp communication.
- Customer sends a message.
- WhatsApp sends the event to your webhook.
- Your backend validates the webhook.
- Your backend checks the customer service window and user state.
- The message is classified: FAQ, order support, sales, complaint, unsafe, or human handoff.
- The backend retrieves relevant data from your CRM, help center, or database.
- The backend sends a controlled prompt to DeepSeek.
- DeepSeek drafts a response.
- Your backend applies safety, policy, and formatting checks.
- The final response is sent through the WhatsApp API.
- If confidence is low, the user is routed to a live agent.
Step 4: Respect templates and the 24-hour window
The most common WhatsApp bot mistake is treating WhatsApp like a free-form chatbot channel. It is not.
Inside the 24-hour customer service window, you can respond to a user message without using a message template. Outside that window, you may only send messages through approved message templates. If your bot needs to re-engage a user after the window closes, design approved templates such as “Your order update is ready” or “A support agent responded to your request,” then continue with free-form support only after the user replies.
Step 5: Add human escalation
A WhatsApp Business API AI chatbot should have live agent handoff by design. The safest implementation is to trigger handoff when:
- The customer asks for a human.
- The user reports an urgent or sensitive issue.
- The model confidence is low.
- The answer depends on private account information.
- The conversation is emotional, legal, medical, or financial.
- The user repeats the same question after a failed answer.
How to Build a DeepSeek Telegram Bot
A DeepSeek Telegram bot is usually faster to create than a WhatsApp bot. Telegram’s official tutorial explains that you obtain a bot token from BotFather by using /newbot, and it warns developers to store the token securely and treat it like a password.
Step 1: Create the bot with BotFather
Open Telegram, message @BotFather, run /newbot, choose a name and username, and save the token in a secret manager or environment variable. Do not hardcode it in your repository.
Step 2: Choose getUpdates or webhooks
Telegram gives you two mutually exclusive ways to receive bot updates:
getUpdates: easier for local testing and simple polling.- Webhooks: better for production because Telegram sends updates to your HTTPS endpoint.
The official Bot API says updates are JSON-serialized and are not kept longer than 24 hours, so production bots should process updates promptly and idempotently.
Step 3: Send user messages to DeepSeek
The basic Telegram Bot API DeepSeek flow is:
- Telegram sends an update to your webhook.
- Your backend extracts
message.textandchat.id. - Your backend sends the text to DeepSeek with a controlled system prompt.
- DeepSeek returns a response.
- Your backend sends the response to Telegram using
sendMessage.
Telegram’s tutorial notes that sending private text generally requires that the user has contacted the bot first, a saved user ID, and a message string of 1–4096 characters.
Minimal Python Webhook Example
The official DeepSeek quick-start docs show the OpenAI-format base URL, environment-variable API key usage, and chat completion call pattern. The following example is intentionally minimal and educational. It does not include production security controls such as webhook secret validation, rate limiting, deduplication, audit logging, or human escalation.
import os
import httpx
from fastapi import FastAPI, Request
app = FastAPI()
TELEGRAM_TOKEN = os.environ["TELEGRAM_BOT_TOKEN"]
DEEPSEEK_API_KEY = os.environ["DEEPSEEK_API_KEY"]
TELEGRAM_API = f"https://api.telegram.org/bot{TELEGRAM_TOKEN}"
DEEPSEEK_API = "https://api.deepseek.com/chat/completions"
SYSTEM_PROMPT = """
You are a concise customer support assistant.
Answer only from approved business information.
If the user asks for sensitive account changes, ask them to contact a human agent.
"""
@app.post("/telegram/webhook")
async def telegram_webhook(request: Request):
update = await request.json()
message = update.get("message") or update.get("edited_message")
if not message or "text" not in message:
return {"ok": True}
chat_id = message["chat"]["id"]
user_text = message["text"][:4000]
async with httpx.AsyncClient(timeout=30) as client:
deepseek_response = await client.post(
DEEPSEEK_API,
headers={
"Authorization": f"Bearer {DEEPSEEK_API_KEY}",
"Content-Type": "application/json",
},
json={
"model": "deepseek-v4-flash",
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user_text},
],
"thinking": {"type": "disabled"},
"stream": False,
},
)
deepseek_response.raise_for_status()
reply = deepseek_response.json()["choices"][0]["message"]["content"][:4096]
await client.post(
f"{TELEGRAM_API}/sendMessage",
json={"chat_id": chat_id, "text": reply},
)
return {"ok": True}For production, add webhook signature or secret-token validation, rate limits, retry handling, message deduplication, monitoring, and a way to escalate to a human or ticketing queue.
No-Code and Low-Code Options
No-code and low-code tools are useful when you need a prototype, internal assistant, or simple workflow. They are less ideal when you need strict compliance, multi-tenant SaaS, heavy customization, advanced observability, or regulated data handling.
Platforms such as n8n, Make, and Albato publish integration pages for DeepSeek with WhatsApp Business Cloud, Telegram, or WhatsApp Business API. These tools can reduce development time by connecting triggers, actions, and API calls visually.
| Tool type | Best for | Pros | Cons | When to avoid it |
|---|---|---|---|---|
| n8n self-hosted | Technical teams wanting visual workflows with control | Flexible, self-hosting option, good for API workflows | Requires DevOps ownership | Avoid if the team cannot maintain infrastructure |
| n8n cloud | Fast automation prototypes | Visual builder, many integrations | Platform costs and limits | Avoid for highly regulated private data without review |
| Make | Business automations and quick prototypes | Easy visual scenarios, many app modules | Complex logic can become hard to maintain | Avoid for large-scale custom routing |
| Albato | Simple no-code app connections | Quick setup for common integrations | Less control than custom code | Avoid when you need deep backend customization |
| Custom HTTP workflow | API-first automations | Works when native modules are missing | Requires API knowledge | Avoid if no one can maintain API failures |
A sensible path is to prototype with no-code, validate the user journey, then rebuild critical production workflows in custom code if security, scale, or compliance requires it.
Custom-Code Approach
A custom backend is the better option when your DeepSeek WhatsApp Business bot or DeepSeek Telegram bot needs advanced control. Choose custom code when you are building:
- A multi-tenant SaaS bot platform.
- A regulated support workflow.
- A bot connected to sensitive CRM data.
- A high-volume customer support bot.
- Advanced routing and live agent handoff.
- Custom memory or RAG.
- Strong observability and SLAs.
- Complex billing, permissions, or user roles.
Recommended stacks include:
| Layer | Practical options |
|---|---|
| API server | Node.js with Express/Fastify, or Python with FastAPI |
| Queue | Redis, BullMQ, Celery, or managed queue services |
| Database | PostgreSQL for users, sessions, conversations, and audit trails |
| Cache | Redis for conversation state and rate limiting |
| Knowledge base | Vector database plus indexed help-center content |
| Hosting | Cloud Functions, AWS Lambda, container services, Kubernetes, or managed PaaS |
| Monitoring | Centralized logs, error tracking, uptime alerts, conversation analytics |
The custom-code approach also lets you implement deterministic business rules before AI generation. For example, order status should come from your database, not from the model’s memory. The model should explain or format the result, not invent it.
DeepSeek Model Selection for Bots
As of June 2026, the official DeepSeek pricing page lists two main API model names: deepseek-v4-flash and deepseek-v4-pro. Both support thinking and non-thinking modes, 1M context length, JSON output, tool calls, and chat prefix completion. DeepSeek’s chat completion API reference also lists deepseek-v4-flash and deepseek-v4-pro as possible model values and describes the thinking switch, streaming, JSON output, and tool call options.
| Bot need | Recommended starting point |
|---|---|
| High-volume FAQ support | deepseek-v4-flash with thinking disabled |
| Lead qualification | deepseek-v4-flash with JSON output |
| Complex troubleshooting | deepseek-v4-pro or thinking enabled |
| Structured CRM actions | JSON output and tool calls |
| Long knowledge-base context | RAG plus model context management |
| Low latency | Non-thinking mode, short prompts, cached context |
| Better reasoning | Thinking mode, stronger model, stricter review |
Avoid using deepseek-chat or deepseek-reasoner as the primary recommendation in new production systems. DeepSeek’s docs state those legacy names are scheduled for deprecation on July 24, 2026, and currently map to deepseek-v4-flash compatibility modes.
Cost Considerations
The basic cost formula is:
Monthly AI cost = input tokens + output tokens + cache behavior + message volume
For a messaging bot, the biggest cost drivers are:
- Number of conversations.
- Average user message length.
- System prompt length.
- Knowledge-base context length.
- Average AI response length.
- Whether context caching applies.
- Whether the bot uses thinking mode.
- Whether failed or retried requests are counted.
- Whether no-code tools charge per operation or scenario.
As of June 2026, DeepSeek lists pricing per 1M tokens. For deepseek-v4-flash, the listed prices are $0.0028 per 1M input tokens on cache hit, $0.14 per 1M input tokens on cache miss, and $0.28 per 1M output tokens. For deepseek-v4-pro, the listed prices are $0.003625 per 1M input tokens on cache hit, $0.435 per 1M input tokens on cache miss, and $0.87 per 1M output tokens. DeepSeek also states prices may vary and recommends checking the pricing page for the most recent information.
| Cost category | WhatsApp Business bot | Telegram bot |
|---|---|---|
| AI API | DeepSeek token usage | DeepSeek token usage |
| Messaging platform | WhatsApp pricing applies by category/rules | No WhatsApp-style template/window pricing model |
| Infrastructure | Backend, database, queue, monitoring | Backend, database, queue, monitoring |
| No-code platform | Operations, scenarios, tasks, or executions | Operations, scenarios, tasks, or executions |
| Support operations | Human handoff, ticketing, QA | Human handoff, moderation, QA |
For WhatsApp, remember that DeepSeek token usage is only one part of the cost. WhatsApp Business Platform pricing is separate and depends on delivered messages, recipient market, and message category such as marketing, utility, authentication, or service. Service messages inside the customer service window may be priced differently, so always check Meta’s current pricing page before forecasting production costs.
Security and Compliance Checklist
Use this checklist before launching a DeepSeek API integration in production.
| Area | Checklist item |
|---|---|
| Secrets | Store DeepSeek API keys and Telegram/WhatsApp tokens in environment variables or a secret manager |
| Webhooks | Validate webhook requests and reject unknown sources |
| Logging | Avoid logging sensitive message content unnecessarily |
| Privacy | Redact PII where possible and publish a privacy policy |
| Consent | Capture WhatsApp opt-in and support opt-out requests |
| WhatsApp rules | Respect templates, customer service windows, and approved use cases |
| Handoff | Provide live agent handoff or a clear support route |
| Data minimization | Send only necessary context to the AI model |
| Sensitive data | Avoid full card numbers, government IDs, health details, and financial identifiers unless properly compliant |
| Abuse prevention | Add rate limiting, spam controls, and blocked-topic handling |
| Reliability | Add retries, fallback messages, and outage handling |
| Monitoring | Track resolution rate, escalation rate, latency, cost, and user satisfaction |
Meta’s policy specifically places responsibility on businesses to secure notices, permissions, and consents, and it prohibits asking people to share sensitive identifiers such as full payment card numbers or personal ID numbers in restricted ways.
Common Use Cases
DeepSeek for WhatsApp Business and Telegram Bots works best when the bot has a defined business purpose.
| Use case | Best channel | Example |
|---|---|---|
| Ecommerce order support | “Where is my order?” | |
| Appointment booking | WhatsApp or Telegram | “Book a consultation next Tuesday.” |
| Lead qualification | “What service are you interested in?” | |
| Internal team assistant | Telegram | “Summarize today’s support tickets.” |
| Knowledge-base bot | Both | “How do I reset my device?” |
| After-sales support | “How do I use this product?” | |
| Community moderation | Telegram | “Answer common group questions.” |
| Multilingual support | Both | “Reply in the customer’s language.” |
The strongest bots combine AI with deterministic systems. For example, a return-policy answer can be generated by DeepSeek, but order eligibility should come from your ecommerce backend.
Common Mistakes to Avoid
The most common mistakes are not model-selection mistakes. They are workflow and compliance mistakes.
| Mistake | Why it matters | Better approach |
|---|---|---|
| Treating WhatsApp like an unrestricted chatbot | WhatsApp has strict business messaging rules | Build around opt-in, templates, windows, and handoff |
| Ignoring message templates | Outbound messages can fail or violate policy | Create approved templates before launch |
| No human handoff | Automation can fail or frustrate users | Add clear escalation paths |
| Using outdated DeepSeek model names | Legacy names are scheduled for deprecation | Use current model names from DeepSeek docs |
| Hardcoding API keys | Tokens can leak | Use secret managers |
| Sending every message to AI | Cost, latency, and risk increase | Route simple events with rules first |
| No fallback | API outages create dead ends | Add graceful failure messages |
| No analytics | You cannot improve what you cannot measure | Track resolution, cost, latency, and escalation |
| Overusing general-purpose AI on WhatsApp | Policy and regulatory risk may apply | Keep business bots purpose-specific and review current terms |
DeepSeek for WhatsApp Business and Telegram Bots: Implementation Plan
Use this 7-step plan to launch safely.
1. Define the use case
Start with one focused goal: ecommerce FAQ, appointment booking, lead qualification, or internal assistant. Do not launch a “do everything” bot first.
2. Choose WhatsApp, Telegram, or both
Choose WhatsApp for official customer conversations. Choose Telegram for fast prototypes, communities, internal tools, and technical audiences. Use both only when the business logic is already stable.
3. Prepare compliance and opt-in
For WhatsApp, document opt-in, opt-out, message templates, human escalation, and privacy notices. For Telegram, document token handling, data retention, and acceptable use.
4. Build a prototype
Use n8n, Make, Albato, or a simple backend to validate the conversation flow. Start with a small set of intents and a fallback path.
5. Connect the DeepSeek API
Use current DeepSeek model names. For everyday support, start with deepseek-v4-flash. For complex reasoning, test deepseek-v4-pro or thinking mode. Keep prompts short, grounded, and business-specific.
6. Add business data and handoff
Connect CRM, order data, ticketing, or knowledge-base content. Add live agent handoff for low-confidence answers, sensitive cases, and user requests.
7. Test, monitor, and scale
Test with real support transcripts, edge cases, multilingual users, broken webhooks, API failures, and policy-sensitive requests. Monitor cost, latency, accuracy, escalation rate, and user satisfaction.
FAQ
Can I use DeepSeek with WhatsApp Business?
Yes. You can connect DeepSeek to WhatsApp Business through the WhatsApp Business Platform or Cloud API, usually via a backend, webhook, or workflow tool. The important part is compliance: opt-in, message templates, the 24-hour customer service window, and human escalation must be designed into the workflow.
Can I create a DeepSeek Telegram bot?
Yes. Create a Telegram bot with BotFather, store the bot token securely, receive updates through webhooks or getUpdates, send user messages to DeepSeek, and return the response with Telegram’s sendMessage method. Telegram’s official docs describe BotFather token creation and update handling.
Is there an official DeepSeek WhatsApp number?
Do not assume that a WhatsApp number or third-party “DeepSeek on WhatsApp” service is official unless DeepSeek itself identifies it as official. For business use, the safer path is to connect your own WhatsApp Business account to the DeepSeek API or use a vetted integration platform.
Is WhatsApp or Telegram better for a DeepSeek bot?
WhatsApp is better for customer support, ecommerce, sales, and official business communication. Telegram is better for quick prototypes, developer tools, internal teams, communities, and flexible bot experiences. For regulated customer workflows, WhatsApp requires more policy planning.
Do I need coding to connect DeepSeek to WhatsApp or Telegram?
Not always. No-code tools such as n8n, Make, and Albato can connect DeepSeek with WhatsApp or Telegram workflows. Coding becomes important when you need custom security, advanced routing, private data controls, high volume, or multi-tenant SaaS logic.
Which DeepSeek model should I use for customer support?
Start with deepseek-v4-flash for high-volume everyday support and test deepseek-v4-pro for complex issues. Use non-thinking mode for faster routine answers and thinking mode for difficult reasoning. Always check the current DeepSeek model and pricing docs before deployment.
Is it compliant to run AI bots on WhatsApp Business?
It can be compliant if the bot follows WhatsApp Business policies, including opt-in, approved message templates where required, the 24-hour customer service window, privacy obligations, and clear escalation paths. General-purpose AI assistant distribution may face additional policy or regulatory complexity, so review current Meta terms before launch.
How much does a DeepSeek messaging bot cost?
The cost depends on DeepSeek token usage, message volume, output length, context size, cache behavior, WhatsApp messaging costs, hosting, and workflow platform fees. As of June 2026, DeepSeek publishes per-1M-token pricing and notes that prices may change, so production forecasts should use current pricing pages, not static estimates.
Can I connect the bot to my CRM or knowledge base?
Yes. A production bot should usually connect to a CRM, ticketing system, ecommerce backend, or knowledge base. The safest pattern is to retrieve verified business data first, then ask DeepSeek to explain, summarize, classify, or format the response.
What is the safest way to launch a production bot?
Start with a narrow use case, approved data sources, strict prompts, webhook validation, API key security, logging controls, human handoff, and clear failure messages. Launch to a small user group first, measure quality, then expand.
Conclusion
DeepSeek for WhatsApp Business and Telegram Bots can be a powerful setup for customer support, lead qualification, internal automation, and knowledge-base assistance. The winning implementation is not just “connect AI to chat.” It is a controlled workflow that respects WhatsApp policies, uses Telegram safely, chooses current DeepSeek models, protects user data, and routes sensitive cases to humans.