Chinese characters garbled into question marks (???) in the Web UI / Playground

When sending requests to the model (e.g., Qwen3-4B-Instruct) via the web interface/playground, the response text displays as a series of question marks (???) instead of the correct Chinese characters.

This issue appears to be a charset encoding mismatch. It seems the backend is correctly returning UTF-8 bytes, but the frontend or the HTTP response headers are incorrectly interpreting them as ASCII or ISO-8859-1.

Steps to reproduce the behavior:

Navigate to the Chat/Playground page.

Enter a simple Chinese prompt, for example: 你好，请介绍一下微信医保支付。

Send the request.

Observe that the output is garbled (e.g., ??????), whereas it should display Chinese characters.

Expected behavior

The UI should correctly render the UTF-8 encoded Chinese characters returned by the model.

Expected Output:

微信支持医保电子凭证...

Actual Output:

??????????????

Screenshots

(If applicable, drag and drop a screenshot of the garbled text here)

Environment (please complete the following information):

OS:​ [e.g., Windows 11 / macOS Sonoma]

Browser:​ [e.g., Chrome 125.0, Edge 124.0]

Version/Branch:​ [e.g., main branch commit hash or Docker image tag]

Additional context & Analysis

I have tested the raw API response using curlor Postman. The raw byte stream contains the correct UTF-8 sequences, which suggests this is not a model training issue but a frontend rendering or HTTP header configuration issue.

Possible causes to investigate:

The HTTP Response Header is missing Content-Type: text/html; charset=utf-8.

The JavaScript frontend is decoding the stream using the wrong encoding (e.g., latin1instead of utf-8).

Database storage collation is set to a non-UTF8 format (less likely for streaming responses).

Test Case:

Sending a simple ASCII prompt works fine, but any Unicode character (Chinese, Emoji) triggers the bug.