FAQ

Frequently asked questions about Alibaba Cloud Model Studio.

Billing

1. What are the unit prices for the models in Alibaba Cloud Model Studio?

   Check in the Model Studio console. For pricing, see Model inference pricing.

2. Are subscription services available?

   Some models support subscription, see Savings plans.

3. Are pay-as-you-go bills settled monthly?

   Bills are calculated per minute and settled monthly.

4. How to query my charges and view details?

   You can go to Expenses and Costs to view the details.

5. How to request an invoice for my expenses?

   On the Overview of Monthly Bill page, click Download Invoice in the Actions column next to the destination account.

6. Does Wan membership covers Model Studio API calls?

   No. Wan membership benefits do not apply to Model Studio API calls because they use separate billing systems.

API/SDK

1. How to view error code information?

   An API call to Model Studio returns a status code that indicates the result of the call. For more information about error codes and their solutions, see Error messages.

2. How to install the SDK?

   Model Studio currently supports SDKs for Java and Python. For more information, see Install the SDK.

3. When calling a function with the assistant API, does it support calling two local functions in sequence?

   a. Calling two functions separately is not currently supported.

   b. You can manually create two separate assistant APIs and handle the response from each one individually.

4. Does the assistant API have memory-related capabilities?

   The memory configuration feature is not currently supported.

Product

1. How to activate the Alibaba Cloud Model Studio service?

   Log on to your Alibaba Cloud account and go to Alibaba Cloud Model Studio (Singapore) or Alibaba Cloud Model Studio (Beijing). On the Homepage or Models page, you can activate the model service. For more information, see Product activation.

   If prompted that real-name verification is incomplete, complete first.

2. How to deactivate the Alibaba Cloud Model Studio service after activation?

   The Model Studio service cannot be deactivated. If you use the API to call models or applications, you can delete your API key on the API-Key (Singapore), API-Key (Virginia), or API-Key (Beijing) page in the console to prevent subsequent calls.

3. How to experience the model service?

   You can experience the service in the Playground (Singapore), Playground (Virginia), or Playground (Beijing).

4. What is the difference between Alibaba Cloud Model Studio and Qwen?

   Alibaba Cloud Model Studio is a model service platform that provides various models, including the Qwen series.

5. How can I implement data isolation to ensure that data from different users is not associated?

   You can use an Alibaba Cloud account to grant different workspace permissions to different RAM users. This practice ensures that data in different workspaces is isolated. For more information, see Workspace permission management.

6. Does Model Studio save the data generated during model calls?

   Alibaba Cloud strictly protects your data privacy and will never use your data for model training. In addition, all data that you transmit when you build applications or train models is encrypted using AES-256 (Advanced Encryption Standard) to ensure data security. When you call a model through Model Studio, how your data is processed and retained depends on how you call the model:

   • Direct calls using an API: Model Studio does not save conversation data. It only desensitizes and records the call status.

   • Use the Model Studio assistant API: To optimize the service experience, Model Studio retains your conversation history. This information currently has no expiration date.

7. How long are conversation records kept in the Model Studio Playground? Is there a limit?

   The console displays up to 100 conversation records with no time limit. Deleting records automatically reveals older ones. Conversations from non-logged-in trial sessions or with inference errors are not saved.

Models

1. How are the parameters of a model stored?

   You can download open source models from the ModelScope community. Their structure is usually defined in a JSON file. You typically need to use open source Python libraries to parse these files. These files contain vector information, and parsing them can help you understand the storage process.

2. How many languages do the Qwen series models support?

   14 languages: Chinese, English, Arabic, Spanish, French, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, and Indonesian.

3. Can the current models connect to structured data, such as MySQL and Hive?

   This is not currently supported. However, this feature is under development, and priority is being given to connectivity with ApsaraDB RDS services.

4. Is the text generation speed of models such as Qwen3 and Qwen-Max fixed for all users, or is there a way to adjust the speed?

   The generation speed is not fixed. It is affected by factors such as the overall load of the current service and your request concurrency.

5. After model throttling is triggered, how long should I generally wait before trying again?

   The waiting time depends on your specific rate limit value, such as requests per second (RPS) or requests per minute (RPM). For example, if your limit is 120 RPM (2 requests per second), and you submit 2 requests consecutively within 0.2 seconds, the 3rd request will be throttled. You will need to wait about 0.8 seconds before you can successfully submit another request.

Model hallucination

1. What is model hallucination?

   Model hallucination refers to the phenomenon where large language models (LLMs) fabricate facts, distort information, or produce logical contradictions when generating content. The output may appear reasonable and fluent but significantly deviates from the input, real-world knowledge, or contextual logic. Note that hallucination differs from factual errors (such as errors caused by outdated training data), subjective opinions, or creative fiction (such as writing a novel when explicitly asked). The core characteristic is confident assertions without basis.

2. How to reduce model hallucination?

   You can reduce model hallucination in the following ways:

   1. Choose a stronger model: Choosing a larger, more advanced model generally reduces hallucination. For example, in the Qwen series, Max-level models outperform Plus-level models, and Plus-level models outperform Turbo-level models.

   2. Prompt engineering: Modifying prompts is a simple yet effective way to reduce model hallucination. For example, in RAG scenarios, add instructions such as "Answer only based on the provided documents. If the information is insufficient, say 'I don't know'"; add "Cite specific data or reports to support your conclusions"; use prompts to break tasks into multiple steps; or assign a rigorous role in the prompt.

   3. RAG (Retrieval-Augmented Generation): Leveraging RAG to provide references for model responses and strictly limiting answers to the scope of retrieved knowledge can significantly reduce hallucination. When building a RAG system, ensure the retrieval system is high-quality, clearly cite information sources, and gracefully handle cases where no relevant information is found.

   4. Plugins/MCP: Use plugins or MCP capabilities to reduce model hallucination. For example, when using an LLM to summarize data from a structured database, you can invoke a database client through plugins or MCP to perform calculations and return the results to the model for summarization. This prevents hallucination that may occur when the model directly handles numerical calculations.

   5. Model parameter tuning : Lower randomness parameters such as temperature, top_k, and top_p to make the output more conservative and avoid bizarre generations, though this may sacrifice creativity. Lowering max_tokens can, in some scenarios, prevent the model from fabricating content after providing key information.

   6. Post-processing verification: After model inference, verify the correctness of the response through subsequent steps. This is typically done by using AI capabilities to check whether the response contains hallucination. This approach increases costs and reduces request processing speed.