Since October 31, 2026, GreenNode MaaS has been serving 8 self-hosted models on GreenNode infrastructure in Vietnam: GLM-5.3, GLM-5.3 Flash, Gemma 4 31B, Qwen3.8-27B, two embedding models (bge-m3, Qwen3-Embedding-8B) and two rerank models (bge-reranker-v2-m3, Qwen3-Reranker-8B). Businesses call these models through an OpenAI-compatible API and pay per token, with no GPUs to operate.
This article introduces each model, suggests which model to use for which workload, and explains how to get started.
What Is a Self-Hosted Model, and How Is It Different from a Third-Party Model?
On GreenNode MaaS, every model is labeled by where requests are processed:
- GreenNode self-host: GreenNode downloads the weights of an open-source model, then deploys and runs it on GreenNode GPUs in Vietnam. Requests and outputs are processed on this infrastructure.
- Third-party: requests are forwarded to the model developer under an official contract and processed on the developer’s infrastructure.
This difference matters for businesses that must comply with rules on processing personal data and transferring it abroad, such as Vietnam’s Personal Data Protection Law, in effect since January 1, 2026. With a self-hosted model, you know exactly where your data is processed. When you need a commercial model that is only available from its developer, you can still use a third-party model on the same platform, with appropriate data. Learn more: Why data processing location matters in the AI era.
If you are comparing Model API providers, How to choose a MaaS platform in Vietnam and APAC compares them in detail across 7 criteria.
8 New Self-Hosted Models on GreenNode MaaS
The eight models fall into five groups by their role in an AI application. GLM-5.3 and GLM-5.3 Flash are the two main models for text generation; the other groups add smaller options and the building blocks for search systems. The official list and availability date of each model are in the launch announcement on GreenNode Docs.
| Group | Model | Developer | Input | Context (per model card) | Best for |
|---|---|---|---|---|---|
| Frontier | GLM-5.3 | Z.ai | Text | 1M tokens | Complex reasoning, coding, agents |
| Workhorse | GLM-5.3 Flash | Z.ai | Text, image, video, file | 1M tokens | Everyday tasks, balancing cost and quality |
| Small | Gemma 4 31B | Google DeepMind | Text, image | 256K tokens | Multilingual tasks, function calling, document understanding |
| Small | Qwen3.8-27B | Qwen | Text, image, video | 262K tokens | Coding, agents, professional tasks |
| Embedding | bge-m3 | BAAI | Text | 8,192 tokens | Multilingual search at low cost |
| Embedding | Qwen3-Embedding-8B | Qwen | Text | 32K tokens | High-accuracy search, long documents |
| Rerank | bge-reranker-v2-m3 | BAAI | Text | Not published | Fast reranking, lightweight model |
| Rerank | Qwen3-Reranker-8B | Qwen | Text | 32K tokens | Accurate reranking, task-specific instructions |
Context is the maximum stated on the developer’s model card. For the configuration served through GreenNode MaaS, see the model list.
Model Groups in Detail
GLM-5.3: a frontier model for complex tasks
GLM-5.3 is Z.ai’s flagship model, released in August 2026, focused on coding and multi-step agent tasks. Reasoning is always on, with low, high or maximum reasoning effort selectable depending on task difficulty. GLM-5.3 currently accepts text input only.
Best for: questions that need multi-step reasoning, writing and fixing code, and agents that automate internal workflows.
GLM-5.3 Flash: the workhorse for everyday tasks
GLM-5.3 Flash is the leaner model in the GLM-5.3 line, introduced by Z.ai as the first natively multimodal model in the GLM-5 series: it accepts text, images, video and files. It is optimized for low cost per task.
Best for: customer service chatbots, internal assistants, summarizing and extracting information from documents, and high-volume daily workloads.
Gemma 4 31B and Qwen3.8-27B: small, fast and cost-efficient
Two dense models of around 30 billion parameters, both under the Apache 2.0 license, suited to simpler tasks that need speed and low cost.
- Gemma 4 31B (Google DeepMind): out-of-the-box support for 35+ languages, pre-trained on 140+ languages; function calling and document understanding, with image input.
- Qwen3.8-27B (Qwen): image and video input, 262K-token context, designed for coding and long-horizon agent tasks, with adjustable reasoning effort.
Best for: text classification, field extraction, batch processing, or as a fallback model when a large model is not needed.
bge-m3 and Qwen3-Embedding-8B: embedding models for search
Embedding models turn text into vectors so you can find relevant documents by meaning. They are the first step of a RAG system.
- bge-m3 (BAAI): 100+ languages, up to 8,192 input tokens, 1,024-dimension vectors, and support for dense, sparse and multi-vector retrieval. A lightweight, widely used option for multilingual search.
- Qwen3-Embedding-8B (Qwen): 100+ languages, 32K-token context, vectors up to 4,096 dimensions with smaller sizes selectable, and task-specific instructions. Best when you need higher accuracy or work with long documents.
bge-reranker-v2-m3 and Qwen3-Reranker-8B: rerank models
Rerank models re-score how relevant each retrieved document is to the question, so the most relevant documents go into the prompt.
- bge-reranker-v2-m3 (BAAI): about 0.6 billion parameters, multilingual, built on bge-m3. Light and fast.
- Qwen3-Reranker-8B (Qwen): 8 billion parameters, 32K-token context, 100+ languages, task-specific instructions. Best when search accuracy is the priority.
The model cards do not list Vietnamese separately among supported languages. Test quality on your own data in your users’ languages before choosing.
Which Model for Which Workload?
The table below is a starting point for testing, not a substitute for measuring on your own data.
| Workload | Start with | Upgrade when needed |
|---|---|---|
| Customer service chatbot | GLM-5.3 Flash | GLM-5.3 for complex questions |
| Internal document assistant (RAG) | bge-m3 + bge-reranker-v2-m3 + GLM-5.3 Flash | Qwen3-Embedding-8B + Qwen3-Reranker-8B |
| High-volume classification and extraction | Gemma 4 31B or Qwen3.8-27B | GLM-5.3 Flash |
| Reading documents with images (invoices, forms) | GLM-5.3 Flash or Qwen3.8-27B | Also test Gemma 4 31B |
| Coding assistance, multi-step agents | GLM-5.3 | Compare with Qwen3.8-27B to optimize cost |
| Semantic search over a large document base | bge-m3 | Qwen3-Embedding-8B for long documents |
A general rule: start with the smallest model that meets your requirements, and move to a larger model only after measuring and finding quality falls short. Larger models usually cost more per token and respond more slowly.
Getting Started with the 8 Self-Hosted Models
- Try them in the Playground. Sign in to GreenNode AI Platform, open the MaaS Playground and run a few real questions against each model for a quick comparison. See also: Test and compare AI models in a no-code Playground.
- Create an API key. New API keys start as pending and must be ACTIVE before use.
- Call the API from your application. The API is OpenAI-compatible, so you can use the existing OpenAI SDK and change only the base URL, API key and model name:
from openai import OpenAI
client = OpenAI(
base_url="https://maas-llm-aiplatform-hcm.api.vngcloud.vn/v1",
api_key="YOUR_GREENNODE_API_KEY",
)
response = client.chat.completions.create(
model="MODEL_ID", # get the exact ID from GET /models
messages=[{"role": "user", "content": "Summarize our return policy in 3 points."}],
)
print(response.choices[0].message.content)- Track cost and rate limits. Usage is billed per input, output and cached token; you can choose prepaid or monthly postpaid billing. The default limit is 10 requests/minute and 14,400 requests/day per account, shared across all models. For higher production limits, contact GreenNode.
Frequently Asked Questions
How is a self-hosted model on GreenNode MaaS different from calling the developer’s API directly?
With a self-hosted model, requests are processed on GreenNode GPUs in Vietnam rather than on the developer’s infrastructure abroad. You also contract with a Vietnamese legal entity and pay in VND.
Do these models work well in Vietnamese?
All of them are multilingual, but the model cards do not publish Vietnamese-specific scores. The most reliable check is to run a set of Vietnamese questions from your real work in the Playground and compare the results.
How is data sent to self-hosted models processed and stored?
Requests are processed on GreenNode infrastructure in Vietnam. [Product/Legal to provide the official answer on log retention and whether data is used for training.]
How are self-hosted models priced?
By input, output and cached tokens, with 1 credit = 1 VND. Per-model prices are shown in the AI Platform Portal. You can choose prepaid credits or monthly postpaid billing.
Can businesses outside Vietnam use these models?
Yes. The API is OpenAI-compatible, so it can be integrated from anywhere. Test latency from your own market and check your local cross-border data transfer rules.
Try the 8 Self-Hosted Models Today
The fastest way to find the right model is to run your own questions in the Playground. When you need higher rate limits or postpaid billing, the GreenNode team will work with you. Contact the GreenNode team now.
