Viewing a Built-in Model in ModelArts Studio (MaaS)

ModelArts Studio provides various open-source models. You can check them on the Model Square. The model details page shows all necessary information. You can choose suitable models for training and inference to incorporate into your enterprise systems.

Prerequisites

You have registered a Huawei account and enabled Huawei Cloud services.

Accessing the Model Square

Log in to the MaaS console and select the target region on the top navigation bar.
In the navigation pane on the left, choose Model Square.

In the Model Filtering area on the Model Square page, filter models by type, context length, advanced features, series, and supported jobs, or search by model name.

For details about model series, see Supported Models. The supported filter criteria may vary depending on the region. The following uses CN-Hong Kong as an example.

**Table 1** Model filtering
Filter Criteria	Description
Type	You can filter models by type, including text generation and image understanding models. If you select multiple model types, the page displays the collection of the selected model types.
Context Length	You can filter models by context length, including 128K, 64K, 32K, 16K, and no more than 8K. If you select multiple context lengths, the page displays the model set of the selected context lengths.
Advanced Capabilities	You can filter models based on their ability to handle function calls or deep reasoning.
Model	You can filter models by series, such as DeepSeek, Qwen 2.5, and DeepSeek Coder. Models supported by each region are different. For details, see Supported Models. If you select multiple models, the page displays the collection of the selected model series.
Supported Job Types	You can filter models by their job types, such as deployment.

Perform the following operations on the target model card on the Model Square page:
A model card displays brief information about a model, such as the model introduction, model type, supported capabilities, context length, and update time.

Figure 1 Model card example
- Hover over the model card to view the operation buttons. Click them as needed.
  Only operations supported by the model are displayed on the model card. These options change based on the specific model. For details about how to deploy a model service, see Deploying a Model Service in ModelArts Studio (MaaS).
- Click the target model card. On the model details page that is displayed, you can view the model introduction, supported versions, functions, and filing information. The capabilities and supported operations may vary depending on the model version.
  - In the upper right corner of the model details page, click Model Deployment or Call to train and run the model. Some options allow you to choose versions.
  - Click the buttons on the right side of the version card to train or use models when needed.
  - The model details page shows the latest version's card information first. For models with multiple versions, older versions are collapsed initially. Click on the left side of a historical version's card to see its details.
    Figure 2 Viewing historical version information
  - If the model involves billing, the version card displays the inference pricing information. Click Switching millions of tokens or Switching to thousands of tokens to adjust the unit of the inference price.
    Figure 3 Inference price

Supported Models

The table below lists the models supported by MaaS. For details about the models, go to the model details page.

**Table 2** Models in the Model Square
Model Series		Type	Use Case	Supported Language	Supported Region	Model Introduction
DeepSeek	DeepSeek-R1	Text generation	Q&A and text generation inference	Chinese and English	CN-Hong Kong and ME-Riyadh	DeepSeek-R1 uses advanced technology for long-context understanding and fast inference. The model supports multimodal interactions and API integrations. It enhances applications like intelligent customer service and data analytics, offering top cost-effectiveness for intelligent upgrades of enterprises.
	DeepSeek-V3	Text generation	Q&A and translation	Chinese and English	CN-Hong Kong	DeepSeek-V3 is a strong MoE language model. It uses a new load balancing method without extra loss and aims for better performance with multi-token predictions.
	DeepSeek-V3.1	Text generation	Q&A	Chinese and English	CN-Hong Kong and ME-Riyadh	DeepSeek-V3.1 is a hybrid model with thinking and non-thinking modes. It matches DeepSeek-R1-0528's performance but offers quicker responses and better tool optimization.
	DeepSeek-V3.2-Exp	Text generation	Q&A	Chinese and English	CN-Hong Kong	V3.2-Exp builds on V3.1-Terminus by introducing DeepSeek sparse attention. It also tests and confirms ways to improve efficiency in long-text training and inference.
	DeepSeek-R1-Distill-Qwen-14B	Text generation	Q&A and text generation inference	Chinese and English	CN-Hong Kong	Qwen-14B is distilled from DeepSeek-R1 outputs and matches the capabilities of OpenAI o1-mini. DeepSeek-R1 performs similarly to OpenAI-o1 in math, coding, and inference tasks.
	DeepSeek-R1-Distill-Qwen-32B	Text generation	Q&A and text generation inference	Chinese and English	CN-Hong Kong	Qwen-32B is distilled from DeepSeek-R1 outputs and matches the capabilities of OpenAI o1-mini. DeepSeek-R1 performs similarly to OpenAI-o1 in math, coding, and inference tasks.
Deepseek-Coder		Text generation	Q&A and text inference	Chinese and English	CN-Hong Kong	DeepSeek Coder includes several code language models. Every model trains from scratch using 2 trillion tokens, with 87% being code and 13% English or Chinese text. It excels in multiple programming languages and performs well across various benchmarks among open-source code models.
Qwen	QwQ	Text generation	Q&A	English	CN-Hong Kong	QwQ is part of the Tongyi series of inference models. Unlike standard instruction-tuning models, QwQ excels at thinking and reasoning, delivering better results on complex tasks.
Qwen 2.5	Qwen 2.5	Text generation	Multilingual processing, mathematical inference, and Q&A	Chinese and English	CN-Hong Kong	Alibaba Cloud's Qwen 2.5 is a new addition to the Qwen series of LLMs. Qwen 2.5 offers various base and instruction-tuned language models, spanning sizes from 0.5 billion to 72 billion parameters.
Qwen 2.5	Qwen2.5-VL	Image understanding	Image understanding and Q&A	Chinese and English	CN-Hong Kong	Qwen-2.5-VL is an open-source multimodal visual language model created by Alibaba Cloud's Qwen team. It excels in visual and language understanding.
Qwen3	Qwen3	Text generation	Q&A	Chinese and English	CN-Hong Kong	The Qwen3 series includes LLMs and multimodal models created by the Qwen team. These models undergo extensive training using vast amounts of language and multimodal data, followed by fine-tuning with top-tier datasets.
Kimi	Kimi-K2	Text generation	Q&A	Chinese and English	CN-Hong Kong	Kimi K2 is a modern MoE language model featuring 32 billion activated parameters and 1 trillion total parameters. Trained using the Muon optimizer, it performs exceptionally well in advanced knowledge, reasoning, and programming tasks while enhancing its agent functionalities.