Ready-to-use AI models for your app: Integrate via API, no training required

Key takeaways

• MaaS turns AI into a standard API call - you handle prompts and parsing, the provider handles training, hosting, and scaling.
• Match the model family (LLM, vision, code, speech) to the task before you compare vendors.
• For apps in Vietnam or Thailand, latency and data sovereignty make a regional MaaS endpoint the safer default.
• You can start with inference alone and layer on fine-tuning, RAG, or agents only when you genuinely need them.

If a product deadline has ever slipped because an AI feature still isn't shipped, you already know the bottleneck. Training a model from scratch demands GPU clusters, labeled datasets, ML engineers, weeks of iteration, and a budget most product teams simply don't have. The good news: you need none of that to add serious AI to your app.

Model as a Service (MaaS) lets you use pre-trained AI models via API - the same way you call a payment gateway or a maps service. Send a request, get a structured response, ship the feature. No infrastructure to own, no MLOps pipeline, no model-versioning overhead on your side. This guide walks through what MaaS requires from you, how to pick the right model, and the one architectural decision teams in Southeast Asia can't afford to skip.

Here's what you actually need to know before picking a provider.

What MaaS does (and doesn't) require from you

At its core, MaaS is a cloud-hosted inference endpoint. The provider trains, hosts, and maintains the model. Your application authenticates with an API key, sends a payload (a prompt, an image, a document), and receives a prediction or generated output. Your job is limited to prompt design, response parsing, and error handling.

What you don't need:

• A GPU cluster or any compute provisioning
• Training data or dataset pipelines
• ML engineering capacity for model development
• Ongoing retraining or model drift monitoring

What you do need:

• A clear definition of the task (classification, generation, extraction, summarization, etc.)
• An API key from a provider whose model fits that task
• Basic HTTP client code in your language of choice

According to a 2024 arXiv survey on MaaS, this model "allows users to access functions of the large model through calling API without the need to train and maintain complex models themselves" - which is exactly the value proposition for teams moving fast.

How to choose the right AI model for your task

Pre-trained models are not interchangeable. Before you integrate, match the model family to the job:

Text and language tasks - LLMs (large language models) like GPT-class models, Claude, or open-source alternatives like Mistral, LLaMA, or DeepSeek handle conversational AI, document summarization, content generation, entity extraction, and question answering. Most expose an OpenAI-compatible chat completions endpoint, which simplifies migration between providers.

Computer vision - Models trained for image classification, object detection, or OCR handle visual inputs. These are critical for use cases like document digitization, retail shelf analysis, or identity verification.

Code generation and analysis - Specialized models like DeepSeek Coder or Code Llama outperform general-purpose LLMs on programming tasks.

Speech and audio - Models like Whisper cover transcription and translation, relevant for call center automation or voice-first interfaces.

If you're evaluating vendors systematically, the Complete Guide for CTOs and AI Leads to Evaluating AI Model-as-a-Service Vendors covers API quality, SDK availability, SLA standards, and compliance criteria worth checking before committing.

The Southeast Asia constraint most guides ignore

Most MaaS guides written for a global audience stop at "get your API key and start calling." That's fine if your users sit in North America or Western Europe. For apps deployed in Vietnam or Thailand, two additional decisions materially affect both performance and legal compliance.

Latency: Routing inference to US-East or EU data centers from Ho Chi Minh City or Bangkok adds 150–300ms of round-trip overhead. For real-time features - chat interfaces, live document processing, interactive assistants - that delay is noticeable to users.

Data sovereignty: Vietnam's data rules were overhauled on 1 January 2026. Decree 13/2023/ND-CP was replaced by the Personal Data Protection Law (91/2025/QH15) and its implementing Decree 356/2025/ND-CP, working alongside the Cybersecurity Law 2018 and the Data Law 2024 (guided by Decree 165/2025/ND-CP). The practical shift for developers: sending prompts that contain personal data to an overseas endpoint counts as a cross-border data transfer, and that transfer now needs a pre-approved impact assessment (CTIA), reviewed by the Ministry of Public Security's cybersecurity unit (A05) before you ship, not audited after the fact. Because the Personal Data Protection Law applies extraterritorially, routing through a foreign provider doesn't move the obligation off your plate; you remain the data controller. For a full breakdown, see cloud compliance in Vietnam.

Practical solution: Run inference on a regional MaaS platform hosted within your jurisdiction, or route requests through a local AI Gateway that enforces data handling policies before forwarding to any upstream model.

How GreenNode's MaaS fits this architecture

GreenNode's serverless AI Model platform hosts a curated library spanning both closed-source and open-source modelmodels, across chat, vision, speech, embedding, and reranking, including OpenAI's GPT-5, Google's Gemini 3, Anthropic's Claude, DeepSeek V4, Alibaba's Qwen 3, and GreenNode's own Vietnamese-tuned GreenMind. All are served from availability zones in Hanoi, Ho Chi Minh City, and Bangkok, so you call the endpoint, the model responds, and no data leaves the region unless you configure it to.

Pricing is per-million tokens, starting around $0.30/M for efficient models like DeepSeek and Qwen, which keeps inference costs predictable at scale. Unlike calling OpenAI or Anthropic directly from a Vietnamese app, you're not routing traffic overseas and not bolting on an offshore data agreement just to satisfy local compliance.

This is also where GreenNode differs from a generic AI proxy. A proxy just forwards the request - anyone can call any model, and the only real limit is the invoice. GreenNode is built so governance rides on every model call, around three questions: who is allowed to call which model for which use case, how much each team is spending in tokens (and when to alert), and whether every request and response is logged enough to satisfy an audit. That's the line between real governance and a pay-and-go free-for-all. The AI Gateway layer is where these rules are enforced centrally - routing, model failover, usage quotas, multi-tenant access control - so each app doesn't reimplement them. If you want the business case behind the architecture, Model as a Service for AI adoption covers why MaaS accelerates time-to-value beyond cost savings.

A minimal integration path

For most developers, the fastest path from zero to a working AI feature looks like this:

1. Define the task precisely: "Summarize customer support tickets to under 50 words" is actionable. "Add AI" is not.
2. Pick a model: Match the model family to the task. For Vietnamese-language understanding, prefer models with documented multilingual performance (GreenMind is specifically benchmarked on Vietnamese datasets including VN-MTEB and VMLU).

3. Authenticate and call the endpoint: Most MaaS providers expose an OpenAI-compatible /chat/completions endpoint, so a working integration is 10–15 lines of Python or Node.js. Because it's OpenAI-compatible, you can reuse the official OpenAI SDK and just point it at the regional endpoint:

   from openai import OpenAI
   
      client = OpenAI(
          base_url="https://maas.greennode.ai/v1",  # regional endpoint — confirm in GreenNode docs
          api_key="YOUR_API_KEY",
      )
   
      resp = client.chat.completions.create(
          model="deepseek-v4",  # pick from the model catalog
          messages=[
              {"role": "system", "content": "Summarize the support ticket in under 50 words."},
              {"role": "user", "content": ticket_text},
          ],
      )
   
      print(resp.choices[0].message.content)

4. Handle errors and rate limits: Implement exponential backoff. Don't assume 100% uptime on the first call.
5. Validate outputs programmatically: For production use, parse and validate model responses against expected schemas rather than passing raw output to downstream systems.

For teams at the earliest stage, AIaaS for Startups walks through how fast-growing companies structure AI adoption without overbuilding infrastructure.

When you might need more than MaaS alone

MaaS handles inference. It doesn't handle everything. A few cases where you'll need to layer in additional capabilities:

• Domain-specific accuracy: If generic model outputs aren't precise enough for a regulated field like legal, medical, or financial, consider fine-tuning. GreenNode's AI Platform supports fine-tuning and full training on the same regional infrastructure.
• Retrieval-augmented generation (RAG): For apps that need to answer questions from private documents or knowledge bases, connecting an LLM to a vector database is standard. This doesn't require training, it augments inference with retrieval.
• Agent workflows: If your app needs multi-step planning, tool use, or autonomous task execution rather than single-turn responses, you're moving from MaaS into AI agent territory - and that's exactly what GreenNode AgentBase is built for: a fully managed platform to deploy, scale, and govern agents (runtime, memory, access control, and an MCP gateway) on the same regional infrastructure as the models you already call through MaaS.

For most app teams in Vietnam and Thailand today, MaaS covers the majority of production AI use cases. The option to fine-tune or retrain is there when you truly need it - and regional infrastructure like GreenNode's means that upgrade path doesn't force you onto a different provider or a different data jurisdiction.

FAQs about MaaS

Do I need to train a model to use AI via API? 

No. With Model as a Service, the provider trains and hosts the model. You authenticate with an API key, send your input, and receive the output - training, scaling, and maintenance stay on the provider's side.

Is MaaS compliant with Vietnam's data protection laws? 

It can be, if inference runs inside the jurisdiction. Since 1 January 2026, personal data is governed by the Personal Data Protection Law (91/2025/QH15) and Decree 356/2025/ND-CP - which replaced Decree 13/2023/ND-CP - alongside the Cybersecurity Law 2018. Using a regional endpoint, such as GreenNode's zones in Hanoi and Ho Chi Minh City, keeps user data in-country and avoids triggering the pre-approved cross-border transfer assessment that overseas endpoints now require.

How much does it cost to use AI models via API? 

Most MaaS providers bill per million tokens. On GreenNode, pricing starts around $0.30/M tokens for efficient models like DeepSeek and Qwen, so cost scales with usage instead of fixed GPU spend.

Can I switch providers easily? 

Usually yes. Because most MaaS platforms expose an OpenAI-compatible /chat/completions endpoint, migrating often means changing a base URL and API key rather than rewriting integration code.

Closed-source or open-source model - which should I pick? 

It depends on the task. Closed-source models (GPT-5, Gemini, Claude) usually lead on general reasoning and multimodal work; open-source models (DeepSeek, Qwen, Llama) give you lower cost and more control. On GreenNode both sit behind the same API, so you can test and switch without re-architecting.

How is this different from a generic AI proxy? 

A proxy just forwards your request - anyone can call any model and the only limit is the bill. GreenNode is built around governance: control over who can call which model for which use case, visibility into each team's usage and spend, and request/response logging for audit - so AI usage stays accountable rather than a free-for-all.

What if I need an agent, not just single model calls? 

That's the point where you move from MaaS to GreenNode AgentBase - a managed platform for deploying and governing AI agents (runtime, memory, access control, MCP gateway) on the same regional infrastructure as the models you call through MaaS.

Do you have models tuned for Vietnamese? 

Yes. GreenMind is GreenNode's own Vietnamese model, evaluated on Vietnamese reasoning tasks like VMLU, alongside a Vietnamese embedding model for semantic search in local applications.

Ready to ship an AI feature?

You can go from idea to a working AI feature in an afternoon: no GPUs, no training pipeline, no compliance detour. Explore GreenNode's Model as a Service to browse the model catalog and test in the no-code playground before you write a line of code.